From 6731c1b19b8bb858f30364ff187f0f73e658ad55 Mon Sep 17 00:00:00 2001
From: lixiaolong <702156524@qq.com>
Date: Tue, 8 Sep 2026 14:26:11 +0800
Subject: [PATCH] =?UTF-8?q?feat(vision):=20=E6=B7=BB=E5=8A=A0=E8=A7=86?=
 =?UTF-8?q?=E8=A7=89=E5=AE=9A=E4=BD=8D=E5=92=8C=E6=A0=87=E5=AE=9A=E5=8A=9F?=
 =?UTF-8?q?=E8=83=BD=E6=94=AF=E6=8C=81?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

- 新增 LLM_CHAT 和 VISION_LOCATE_TARGET 动作枚举类型
- 实现 RGBD 相机快照捕获功能，包括深度数据处理
- 集成 OpenCV 库用于视觉标定和标记检测
- 添加视觉定位操作服务和数学计算模块
- 在应用菜单中新增视觉标定管理界面
- 优化 TTS 服务调用和流式处理机制
- 添加闲聊功能支持和默认系统提示配置
- 更新应用配置文件以支持新的视觉功能模块
---
 .gitignore                                    |   1 +
 .../web/controller/test/TeFlowController.java |  87 +++
 .../test/VisionCalibrationController.java     | 133 ++++
 .../src/main/resources/application-cangan.yml |  12 +-
 .../src/main/resources/application.yml        |   2 +-
 .../model/camera/EdgeCameraRgbdSnapshot.java  |  23 +
 .../client/service/EdgeCameraService.java     |   3 +
 .../service/impl/EdgeCameraServiceImpl.java   |  50 ++
 .../llm/service/impl/LLMAiTtsServiceImpl.java |  40 +-
 .../com/cmvr/llm/util/LlmChatService.java     |  13 +-
 cmvr-iot-test/pom.xml                         |   5 +
 .../cmvr/test/constant/LLMChatConstants.java  |  45 ++
 .../java/com/cmvr/test/enums/ActionEnum.java  |   2 +
 .../runtime/operator/edge/VisionGeometry.java | 118 ++++
 .../edge/VisionLocateOperateService.java      | 643 ++++++++++++++++++
 .../operator/llm/LLMChatOperateService.java   |  71 ++
 .../llm/LLMMediaAnalysisOperateService.java   |  26 +-
 .../service/FlowActionExecutorService.java    |  10 +
 .../domain/VisionCalibrationProfile.java      |  37 +
 .../domain/VisionCalibrationRecord.java       |  32 +
 .../VisionCalibrationProfileMapper.java       |   7 +
 .../mapper/VisionCalibrationRecordMapper.java |   7 +
 .../test/vision/marker/MarkerObservation.java |   8 +
 .../vision/marker/SquareMarkerDetector.java   | 287 ++++++++
 .../test/vision/math/CalibrationResult.java   |   5 +
 .../test/vision/math/CalibrationSample.java   |   4 +
 .../vision/math/VisionCalibrationMath.java    | 355 ++++++++++
 .../model/VisionCalibrationRequests.java      |  46 ++
 .../model/VisionCalibrationSessionView.java   |  26 +
 .../VisionCalibrationProfileService.java      | 190 ++++++
 .../VisionCalibrationRecordService.java       |  11 +
 .../VisionCalibrationSessionService.java      | 563 +++++++++++++++
 pom.xml                                       |   7 +
 sql/resource_center_menu.sql                  |  10 +-
 sql/vision_calibration.sql                    |  73 ++
 35 files changed, 2929 insertions(+), 23 deletions(-)
 create mode 100644 cmvr-iot-admin/src/main/java/com/cmvr/web/controller/test/VisionCalibrationController.java
 create mode 100644 cmvr-iot-api/cmvr-iot-edge/cmvr-iot-grpc-client/src/main/java/com/cmvr/edge/client/model/camera/EdgeCameraRgbdSnapshot.java
 create mode 100644 cmvr-iot-test/src/main/java/com/cmvr/test/constant/LLMChatConstants.java
 create mode 100644 cmvr-iot-test/src/main/java/com/cmvr/test/flow/runtime/operator/edge/VisionGeometry.java
 create mode 100644 cmvr-iot-test/src/main/java/com/cmvr/test/flow/runtime/operator/edge/VisionLocateOperateService.java
 create mode 100644 cmvr-iot-test/src/main/java/com/cmvr/test/flow/runtime/operator/llm/LLMChatOperateService.java
 create mode 100644 cmvr-iot-test/src/main/java/com/cmvr/test/vision/domain/VisionCalibrationProfile.java
 create mode 100644 cmvr-iot-test/src/main/java/com/cmvr/test/vision/domain/VisionCalibrationRecord.java
 create mode 100644 cmvr-iot-test/src/main/java/com/cmvr/test/vision/mapper/VisionCalibrationProfileMapper.java
 create mode 100644 cmvr-iot-test/src/main/java/com/cmvr/test/vision/mapper/VisionCalibrationRecordMapper.java
 create mode 100644 cmvr-iot-test/src/main/java/com/cmvr/test/vision/marker/MarkerObservation.java
 create mode 100644 cmvr-iot-test/src/main/java/com/cmvr/test/vision/marker/SquareMarkerDetector.java
 create mode 100644 cmvr-iot-test/src/main/java/com/cmvr/test/vision/math/CalibrationResult.java
 create mode 100644 cmvr-iot-test/src/main/java/com/cmvr/test/vision/math/CalibrationSample.java
 create mode 100644 cmvr-iot-test/src/main/java/com/cmvr/test/vision/math/VisionCalibrationMath.java
 create mode 100644 cmvr-iot-test/src/main/java/com/cmvr/test/vision/model/VisionCalibrationRequests.java
 create mode 100644 cmvr-iot-test/src/main/java/com/cmvr/test/vision/model/VisionCalibrationSessionView.java
 create mode 100644 cmvr-iot-test/src/main/java/com/cmvr/test/vision/service/VisionCalibrationProfileService.java
 create mode 100644 cmvr-iot-test/src/main/java/com/cmvr/test/vision/service/VisionCalibrationRecordService.java
 create mode 100644 cmvr-iot-test/src/main/java/com/cmvr/test/vision/service/VisionCalibrationSessionService.java
 create mode 100644 sql/vision_calibration.sql

diff --git a/.gitignore b/.gitignore
index 2163da6..eeb64bc 100644
--- a/.gitignore
+++ b/.gitignore
@@ -49,3 +49,4 @@ nbdist/
 # Local verification tests
 /cmvr-iot-test/src/test/java/com/cmvr/test/flow/runtime/operator/edge/EdgeAgvOperateServiceTest.java
 /cmvr-iot-test/src/test/java/com/cmvr/test/flow/runtime/operator/edge/EdgeArmOperateServiceTest.java
+/cmvr-iot-test/src/test/java/com/cmvr/test/vision/math/VisionCalibrationMathTest.java
diff --git a/cmvr-iot-admin/src/main/java/com/cmvr/web/controller/test/TeFlowController.java b/cmvr-iot-admin/src/main/java/com/cmvr/web/controller/test/TeFlowController.java
index 4247d60..0e3a043 100644
--- a/cmvr-iot-admin/src/main/java/com/cmvr/web/controller/test/TeFlowController.java
+++ b/cmvr-iot-admin/src/main/java/com/cmvr/web/controller/test/TeFlowController.java
@@ -17,6 +17,7 @@ import com.cmvr.test.flow.runtime.engine.FlowTaskRuntimeService;
 import com.cmvr.test.flow.runtime.subflow.FlowSubFlowDefinitionService;
 import com.cmvr.test.flow.runtime.operator.edge.ti.TiTouchOperateService;
 import com.cmvr.test.model.vo.FlowActionRequestVO;
+import com.cmvr.test.constant.LLMChatConstants;
 import com.cmvr.test.model.vo.TeDetectItemDeployFlowVO;
 import com.cmvr.test.model.vo.TeFlowPublishVO;
 import com.cmvr.test.model.vo.TeFlowVersionActionVO;
@@ -41,6 +42,7 @@ import org.springframework.web.bind.annotation.RequestParam;
 import org.springframework.web.bind.annotation.RestController;
 import org.springframework.web.bind.annotation.PutMapping;
 import org.springframework.beans.factory.annotation.Qualifier;
+import org.springframework.beans.factory.annotation.Value;
 import org.springframework.scheduling.concurrent.ThreadPoolTaskExecutor;
 
 import jakarta.annotation.PreDestroy;
@@ -104,6 +106,18 @@ public class TeFlowController extends BaseController {
      */
     private volatile Future<JSONObject> currentAiAgentFuture;
 
+    /** 闲聊接口使用独立任务，避免影响原商道智能体接口。 */
+    private final ExecutorService chatExecutor = Executors.newSingleThreadExecutor(runnable -> {
+        Thread thread = new Thread(runnable, "flow-llm-chat-query");
+        thread.setDaemon(true);
+        return thread;
+    });
+
+    private volatile Future<String> currentChatFuture;
+
+    @Value("${CMVR_CHAT_TTS_STOP_URL:http://192.168.0.148:8080/tts/stop}")
+    private String chatTtsStopUrl;
+
     @ApiOperation("流程发布")
     @PostMapping("/publish")
     public AjaxResult publish(@Valid @RequestBody TeFlowPublishVO request) {
@@ -367,6 +381,56 @@ public class TeFlowController extends BaseController {
         return AjaxResult.ok("ok");
     }
 
+    /**
+     * 调用工作流闲聊组件。模型响应由组件流式接收，开启 TTS 后按句即时播放，
+     * HTTP 接口在组件完成后返回完整文本。
+     */
+    @GetMapping("/llm/chat/query")
+    @ApiOperation("调用闲聊组件")
+    public AjaxResult queryChat(String text,
+                                @RequestParam(defaultValue = "true") boolean invokeTts) {
+        if (StringUtils.isBlank(text)) {
+            return AjaxResult.error("text不能为空");
+        }
+
+        stopCurrentChatFuture(true);
+        JSONObject payload = new JSONObject();
+        payload.put("text", text.trim());
+        payload.put("promptMode", "CUSTOM");
+        payload.put("prompt", LLMChatConstants.DEFAULT_SYSTEM_PROMPT);
+        payload.put("invokeTts", invokeTts);
+
+        FlowActionRequestVO request = new FlowActionRequestVO();
+        request.setAction("LLM_CHAT");
+        request.setPayload(payload);
+
+        Future<String> future = chatExecutor.submit(() -> flowActionExecutorService.actionExecute(request));
+        currentChatFuture = future;
+        try {
+            return AjaxResult.ok(com.alibaba.fastjson2.JSON.parseObject(future.get()));
+        } catch (CancellationException e) {
+            return AjaxResult.error("闲聊调用已停止");
+        } catch (InterruptedException e) {
+            future.cancel(true);
+            Thread.currentThread().interrupt();
+            return AjaxResult.error("闲聊调用被中断");
+        } catch (ExecutionException e) {
+            Throwable cause = e.getCause() == null ? e : e.getCause();
+            return AjaxResult.error("闲聊调用失败：" + cause.getMessage());
+        } finally {
+            if (currentChatFuture == future) {
+                currentChatFuture = null;
+            }
+        }
+    }
+
+    @PostMapping("/llm/chat/stop")
+    @ApiOperation("停止闲聊组件和TTS")
+    public AjaxResult stopChat() {
+        stopCurrentChatFuture(true);
+        return AjaxResult.ok("ok");
+    }
+
     @GetMapping("/testrun")
     @ApiOperation("测试接口")
     public AjaxResult testrun(@RequestParam("imageUrl") String imageUrl,
@@ -390,6 +454,27 @@ public class TeFlowController extends BaseController {
         }
     }
 
+    private void stopCurrentChatFuture(boolean stopTts) {
+        Future<String> future = currentChatFuture;
+        if (future != null && !future.isDone()) {
+            future.cancel(true);
+        }
+        currentChatFuture = null;
+        if (stopTts) {
+            stopChatTts();
+        }
+    }
+
+    private String stopChatTts() {
+        try (HttpResponse response = HttpRequest.post(chatTtsStopUrl)
+                .timeout(3000)
+                .execute()) {
+            return response.body();
+        } catch (Exception e) {
+            return "调用闲聊TTS停止接口失败：" + e.getMessage();
+        }
+    }
+
     /**
      * 调用本地 TTS 停止接口。
      */
@@ -409,6 +494,8 @@ public class TeFlowController extends BaseController {
     @PreDestroy
     public void destroy() {
         stopCurrentAiAgentFuture(true);
+        stopCurrentChatFuture(true);
         aiAgentExecutor.shutdownNow();
+        chatExecutor.shutdownNow();
     }
 }
diff --git a/cmvr-iot-admin/src/main/java/com/cmvr/web/controller/test/VisionCalibrationController.java b/cmvr-iot-admin/src/main/java/com/cmvr/web/controller/test/VisionCalibrationController.java
new file mode 100644
index 0000000..047cfcc
--- /dev/null
+++ b/cmvr-iot-admin/src/main/java/com/cmvr/web/controller/test/VisionCalibrationController.java
@@ -0,0 +1,133 @@
+package com.cmvr.web.controller.test;
+
+import com.cmvr.common.core.controller.BaseController;
+import com.cmvr.common.core.domain.AjaxResult;
+import com.cmvr.test.vision.domain.VisionCalibrationProfile;
+import com.cmvr.test.vision.model.VisionCalibrationRequests;
+import com.cmvr.test.vision.service.VisionCalibrationProfileService;
+import com.cmvr.test.vision.service.VisionCalibrationRecordService;
+import com.cmvr.test.vision.service.VisionCalibrationSessionService;
+import io.swagger.annotations.Api;
+import io.swagger.annotations.ApiOperation;
+import lombok.RequiredArgsConstructor;
+import org.springframework.web.bind.annotation.DeleteMapping;
+import org.springframework.web.bind.annotation.GetMapping;
+import org.springframework.web.bind.annotation.PathVariable;
+import org.springframework.web.bind.annotation.PostMapping;
+import org.springframework.web.bind.annotation.PutMapping;
+import org.springframework.web.bind.annotation.RequestBody;
+import org.springframework.web.bind.annotation.RequestMapping;
+import org.springframework.web.bind.annotation.RequestParam;
+import org.springframework.web.bind.annotation.RestController;
+
+@Api(tags = "视觉标定管理")
+@RestController
+@RequestMapping("/vision/calibration")
+@RequiredArgsConstructor
+public class VisionCalibrationController extends BaseController {
+    private final VisionCalibrationProfileService profileService;
+    private final VisionCalibrationRecordService recordService;
+    private final VisionCalibrationSessionService sessionService;
+
+    @ApiOperation("查询标定方案")
+    @GetMapping("/profiles")
+    public AjaxResult profiles(@RequestParam(required = false) String profileType,
+                               @RequestParam(required = false) String robotId,
+                               @RequestParam(required = false) String keyword) {
+        return AjaxResult.ok(profileService.list(profileType, robotId, keyword));
+    }
+
+    @ApiOperation("查询标定方案详情")
+    @GetMapping("/profiles/{id}")
+    public AjaxResult profile(@PathVariable String id) {
+        return AjaxResult.ok(profileService.require(id));
+    }
+
+    @ApiOperation("获取当前启用的手眼标定")
+    @GetMapping("/profiles/active/hand-eye")
+    public AjaxResult activeHandEye(String robotId, String armDeviceId, String cameraDeviceId) {
+        return AjaxResult.ok(profileService.activeHandEye(robotId, armDeviceId, cameraDeviceId));
+    }
+
+    @ApiOperation("获取当前启用的工具TCP")
+    @GetMapping("/profiles/active/tool")
+    public AjaxResult activeTool(String robotId, String armDeviceId, String toolCode) {
+        return AjaxResult.ok(profileService.activeTool(robotId, armDeviceId, toolCode));
+    }
+
+    @ApiOperation("保存六自由度人工修正")
+    @PutMapping("/profiles/{id}/correction")
+    public AjaxResult correction(@PathVariable String id,
+                                 @RequestBody VisionCalibrationRequests.Correction request) {
+        return AjaxResult.ok(profileService.updateCorrection(id, request, getUsername()));
+    }
+
+    @ApiOperation("启用已验证方案")
+    @PutMapping("/profiles/{id}/activate")
+    public AjaxResult activate(@PathVariable String id,
+                               @RequestBody(required = false) VisionCalibrationRequests.Activate request) {
+        boolean forced = request != null && Boolean.TRUE.equals(request.getForced());
+        return AjaxResult.ok(profileService.activate(id, forced, getUsername()));
+    }
+
+    @ApiOperation("将方案标记为失效")
+    @DeleteMapping("/profiles/{id}")
+    public AjaxResult invalidate(@PathVariable String id) {
+        profileService.invalidate(id, getUsername());
+        return AjaxResult.success();
+    }
+
+    @ApiOperation("查询标定执行记录")
+    @GetMapping("/records")
+    public AjaxResult records() {
+        return AjaxResult.ok(recordService.list());
+    }
+
+    @ApiOperation("开始全自动手眼标定")
+    @PostMapping("/sessions/hand-eye")
+    public AjaxResult startHandEye(@RequestBody VisionCalibrationRequests.HandEyeStart request) {
+        return AjaxResult.ok(sessionService.startHandEye(request, getUsername()));
+    }
+
+    @ApiOperation("开始工具TCP标定")
+    @PostMapping("/sessions/tool")
+    public AjaxResult startTool(@RequestBody VisionCalibrationRequests.ToolStart request) {
+        return AjaxResult.ok(sessionService.startTool(request, getUsername()));
+    }
+
+    @ApiOperation("记录工具定点旋转姿态")
+    @PostMapping("/sessions/{id}/tool/pivot")
+    public AjaxResult recordToolPivot(@PathVariable String id) {
+        return AjaxResult.ok(sessionService.recordToolPivot(id));
+    }
+
+    @ApiOperation("记录工具安全方向并完成求解")
+    @PostMapping("/sessions/{id}/tool/direction")
+    public AjaxResult finishTool(@PathVariable String id) {
+        return AjaxResult.ok(sessionService.recordToolDirectionAndFinish(id));
+    }
+
+    @ApiOperation("开始人工修正后的快速验证")
+    @PostMapping("/profiles/{id}/verify")
+    public AjaxResult verify(@PathVariable String id) {
+        return AjaxResult.ok(sessionService.startVerification(id, getUsername()));
+    }
+
+    @ApiOperation("标定会话心跳")
+    @PostMapping("/sessions/{id}/heartbeat")
+    public AjaxResult heartbeat(@PathVariable String id) {
+        return AjaxResult.ok(sessionService.heartbeat(id));
+    }
+
+    @ApiOperation("查询标定会话")
+    @GetMapping("/sessions/{id}")
+    public AjaxResult session(@PathVariable String id) {
+        return AjaxResult.ok(sessionService.status(id));
+    }
+
+    @ApiOperation("停止标定会话并停止机械臂")
+    @PostMapping("/sessions/{id}/stop")
+    public AjaxResult stop(@PathVariable String id) {
+        return AjaxResult.ok(sessionService.stop(id));
+    }
+}
diff --git a/cmvr-iot-admin/src/main/resources/application-cangan.yml b/cmvr-iot-admin/src/main/resources/application-cangan.yml
index a482a69..0a7e9bc 100644
--- a/cmvr-iot-admin/src/main/resources/application-cangan.yml
+++ b/cmvr-iot-admin/src/main/resources/application-cangan.yml
@@ -140,9 +140,9 @@ media-analysis:
   read-timeout-ms: 600000
 
 # Isolate the resource-center development instance from the currently running robot gateway.
-cmvr:
-  resource:
-    robot-state:
-      enabled: false
-  quic:
-    enabled: false
+#cmvr:
+#  resource:
+#    robot-state:
+#      enabled: true
+#  quic:
+#    enabled: true
diff --git a/cmvr-iot-admin/src/main/resources/application.yml b/cmvr-iot-admin/src/main/resources/application.yml
index 45de4dc..c2db5cb 100644
--- a/cmvr-iot-admin/src/main/resources/application.yml
+++ b/cmvr-iot-admin/src/main/resources/application.yml
@@ -60,7 +60,7 @@ spring:
     # 国际化资源文件路径
     basename: i18n/messages
   profiles:
-    active: aima
+    active: test
   # 文件上传
   servlet:
     multipart:
diff --git a/cmvr-iot-api/cmvr-iot-edge/cmvr-iot-grpc-client/src/main/java/com/cmvr/edge/client/model/camera/EdgeCameraRgbdSnapshot.java b/cmvr-iot-api/cmvr-iot-edge/cmvr-iot-grpc-client/src/main/java/com/cmvr/edge/client/model/camera/EdgeCameraRgbdSnapshot.java
new file mode 100644
index 0000000..9d81510
--- /dev/null
+++ b/cmvr-iot-api/cmvr-iot-edge/cmvr-iot-grpc-client/src/main/java/com/cmvr/edge/client/model/camera/EdgeCameraRgbdSnapshot.java
@@ -0,0 +1,23 @@
+package com.cmvr.edge.client.model.camera;
+
+import cmvr.api.CameraCommand;
+import lombok.Builder;
+import lombok.Getter;
+
+/** A same-capture RGB-D frame kept in memory for deterministic depth lookup. */
+@Getter
+@Builder
+public class EdgeCameraRgbdSnapshot {
+    private final String colorImageUrl;
+    private final byte[] depthData;
+    private final int width;
+    private final int height;
+    private final CameraCommand.FrameData.FrameType depthType;
+    private final double depthUnitScale;
+    private final double fx;
+    private final double fy;
+    private final double cx;
+    private final double cy;
+    private final long captureTimestamp;
+    private final long sourceFrameNumber;
+}
diff --git a/cmvr-iot-api/cmvr-iot-edge/cmvr-iot-grpc-client/src/main/java/com/cmvr/edge/client/service/EdgeCameraService.java b/cmvr-iot-api/cmvr-iot-edge/cmvr-iot-grpc-client/src/main/java/com/cmvr/edge/client/service/EdgeCameraService.java
index 7aaafeb..83fd87d 100644
--- a/cmvr-iot-api/cmvr-iot-edge/cmvr-iot-grpc-client/src/main/java/com/cmvr/edge/client/service/EdgeCameraService.java
+++ b/cmvr-iot-api/cmvr-iot-edge/cmvr-iot-grpc-client/src/main/java/com/cmvr/edge/client/service/EdgeCameraService.java
@@ -3,6 +3,7 @@ package com.cmvr.edge.client.service;
 import cmvr.api.CameraCommand;
 import com.cmvr.edge.client.model.EdgeCommonVO;
 import com.cmvr.edge.client.model.camera.EdgeCameraPtzVO;
+import com.cmvr.edge.client.model.camera.EdgeCameraRgbdSnapshot;
 import io.grpc.stub.StreamObserver;
 
 import java.io.File;
@@ -64,6 +65,8 @@ public interface EdgeCameraService {
      */
     public String getRGBDImages(EdgeCommonVO edgeCommonVO);
 
+    EdgeCameraRgbdSnapshot captureRgbd(EdgeCommonVO edgeCommonVO);
+
     String controlPtz(EdgeCameraPtzVO ptzVO);
 
     StreamObserver<CameraCommand.GetRGBImageStreamCommand.Request> getRGBImageStream(EdgeCommonVO edgeCommonVO);
diff --git a/cmvr-iot-api/cmvr-iot-edge/cmvr-iot-grpc-client/src/main/java/com/cmvr/edge/client/service/impl/EdgeCameraServiceImpl.java b/cmvr-iot-api/cmvr-iot-edge/cmvr-iot-grpc-client/src/main/java/com/cmvr/edge/client/service/impl/EdgeCameraServiceImpl.java
index 86ab433..f2866f9 100644
--- a/cmvr-iot-api/cmvr-iot-edge/cmvr-iot-grpc-client/src/main/java/com/cmvr/edge/client/service/impl/EdgeCameraServiceImpl.java
+++ b/cmvr-iot-api/cmvr-iot-edge/cmvr-iot-grpc-client/src/main/java/com/cmvr/edge/client/service/impl/EdgeCameraServiceImpl.java
@@ -13,6 +13,7 @@ import com.cmvr.edge.client.manage.GrpcServiceManager;
 import com.cmvr.edge.client.manage.GrpcStreamManager;
 import com.cmvr.edge.client.model.EdgeCommonVO;
 import com.cmvr.edge.client.model.camera.EdgeCameraPtzVO;
+import com.cmvr.edge.client.model.camera.EdgeCameraRgbdSnapshot;
 import com.cmvr.edge.client.service.EdgeCameraService;
 import com.cmvr.edge.client.service.EdgeStreamService;
 import com.cmvr.edge.client.utils.EdgeCommonUtil;
@@ -323,6 +324,55 @@ public class EdgeCameraServiceImpl implements EdgeCameraService, EdgeStreamServi
 //        return JSON.toJSONString(result);
     }
 
+    @Override
+    public EdgeCameraRgbdSnapshot captureRgbd(EdgeCommonVO edgeCommonVO) {
+        if (!checkStatus(edgeCommonVO)) {
+            start(edgeCommonVO);
+        }
+        CameraServiceGrpc.CameraServiceBlockingStub stub = grpcServiceManager.getGrpcClient(
+                edgeCommonVO.getRobotId(), CameraServiceGrpc.CameraServiceBlockingStub.class);
+        CameraCommand.GetRGBDImagesCommand.Request request = CameraCommand.GetRGBDImagesCommand.Request.newBuilder()
+                .setHeader(EdgeCommonUtil.buildRequest(edgeCommonVO.getDeviceId()))
+                .build();
+        CameraCommand.GetRGBDImagesCommand.Feedback feedback = executeGrpcCall(() -> stub.getRGBDImages(request));
+        requireSuccess(feedback.getHeader(), "获取RGBD图像");
+        if (!feedback.hasColorFrame() || !feedback.hasDepthFrame()) {
+            throw new GlobalException("RGBD响应缺少彩色图或深度图");
+        }
+        CameraCommand.FrameData color = feedback.getColorFrame();
+        CameraCommand.FrameData depth = feedback.getDepthFrame();
+        if (color.getWidth() <= 0 || color.getHeight() <= 0
+                || depth.getWidth() != color.getWidth() || depth.getHeight() != color.getHeight()) {
+            throw new GlobalException("RGBD图像尺寸不一致");
+        }
+        String colorPath = saveColorImage(color);
+        String colorUrl;
+        try {
+            colorUrl = sysFileService.uploadFile(convertFileToMultipartFile(colorPath), FileType.IMAGE.code());
+        } catch (IOException exception) {
+            throw new GlobalException("彩色图上传失败: " + exception.getMessage());
+        } finally {
+            deleteTempFile(colorPath);
+        }
+        if (StrUtil.isBlank(colorUrl)) {
+            throw new GlobalException("彩色图上传失败: 未返回地址");
+        }
+        return EdgeCameraRgbdSnapshot.builder()
+                .colorImageUrl(colorUrl)
+                .depthData(depth.getData().toByteArray())
+                .width(depth.getWidth())
+                .height(depth.getHeight())
+                .depthType(depth.getType())
+                .depthUnitScale(depth.getType() == CameraCommand.FrameData.FrameType.F32C1 ? 1.0 : 0.001)
+                .fx(feedback.getIntrinsics().getFx())
+                .fy(feedback.getIntrinsics().getFy())
+                .cx(feedback.getIntrinsics().getCx())
+                .cy(feedback.getIntrinsics().getCy())
+                .captureTimestamp(depth.getCaptureUtcNs())
+                .sourceFrameNumber(depth.getSourceFrameNumber())
+                .build();
+    }
+
     @Override
     public String controlPtz(EdgeCameraPtzVO ptzVO) {
         CameraCommand.ControlPtzCommand.Command command;
diff --git a/cmvr-iot-api/cmvr-iot-llm/src/main/java/com/cmvr/llm/service/impl/LLMAiTtsServiceImpl.java b/cmvr-iot-api/cmvr-iot-llm/src/main/java/com/cmvr/llm/service/impl/LLMAiTtsServiceImpl.java
index b49f08b..754d45c 100644
--- a/cmvr-iot-api/cmvr-iot-llm/src/main/java/com/cmvr/llm/service/impl/LLMAiTtsServiceImpl.java
+++ b/cmvr-iot-api/cmvr-iot-llm/src/main/java/com/cmvr/llm/service/impl/LLMAiTtsServiceImpl.java
@@ -1,18 +1,19 @@
 package com.cmvr.llm.service.impl;
 
 import com.alibaba.fastjson2.JSONObject;
-import com.cmvr.common.utils.http.CallAPIUtil;
-import com.cmvr.llm.service.LLMAiAgentPlatformService;
+import cn.hutool.core.util.StrUtil;
+import cn.hutool.http.ContentType;
+import cn.hutool.http.HttpRequest;
+import cn.hutool.http.HttpResponse;
 import com.cmvr.llm.service.LLMAiTtsService;
-import com.cmvr.llm.util.LlmChatService;
+import com.cmvr.common.exception.GlobalException;
 import lombok.RequiredArgsConstructor;
+import lombok.extern.slf4j.Slf4j;
 import org.springframework.stereotype.Service;
 
-import java.util.HashMap;
-import java.util.Map;
-
 @Service
 @RequiredArgsConstructor
+@Slf4j
 public class LLMAiTtsServiceImpl implements LLMAiTtsService {
     @Override
     public JSONObject play(String url, String text, String voice, String speed, String volume) {
@@ -33,8 +34,29 @@ public class LLMAiTtsServiceImpl implements LLMAiTtsService {
 //            throw new RuntimeException(e);
 //        }
         // 通过post调用tts接口
-        String result = CallAPIUtil.doPostJson(url, new HashMap<>(), new JSONObject().fluentPut("text", text).fluentPut("voice", voice).fluentPut("speed", speed).fluentPut("volume", volume));
-
-        return new JSONObject().fluentPut("result", result);
+        if (StrUtil.isBlank(url)) {
+            throw new GlobalException("TTS播放地址不能为空");
+        }
+        JSONObject body = new JSONObject()
+                .fluentPut("text", text)
+                .fluentPut("voice", voice)
+                .fluentPut("speed", speed)
+                .fluentPut("volume", volume);
+        try (HttpResponse response = HttpRequest.post(url)
+                .contentType(ContentType.JSON.toString())
+                .body(body.toJSONString())
+                .timeout(300000)
+                .execute()) {
+            String result = response.body();
+            if (!response.isOk()) {
+                throw new GlobalException("TTS播放调用失败(" + response.getStatus() + "): " + result);
+            }
+            log.debug("TTS播放调用成功，地址: {}，文本长度: {}", url, text == null ? 0 : text.length());
+            return new JSONObject().fluentPut("result", result);
+        } catch (GlobalException exception) {
+            throw exception;
+        } catch (Exception exception) {
+            throw new GlobalException("TTS播放服务不可用: " + exception.getMessage());
+        }
     }
 }
diff --git a/cmvr-iot-api/cmvr-iot-llm/src/main/java/com/cmvr/llm/util/LlmChatService.java b/cmvr-iot-api/cmvr-iot-llm/src/main/java/com/cmvr/llm/util/LlmChatService.java
index 2fc2cb8..2cb74b9 100644
--- a/cmvr-iot-api/cmvr-iot-llm/src/main/java/com/cmvr/llm/util/LlmChatService.java
+++ b/cmvr-iot-api/cmvr-iot-llm/src/main/java/com/cmvr/llm/util/LlmChatService.java
@@ -273,6 +273,8 @@ public class LlmChatService {
                     // 流正常结束
                     fullContent.append(localContent);
                     finalMessageIdRef.set(finalMessageId);
+                    // Some agents close SSE without an explicit end event.
+                    streamFinished.set(true);
 
                 } catch (IOException e) {
                     errorRef.set(e);
@@ -292,6 +294,7 @@ public class LlmChatService {
                 if (!playTts) return;
 
                 sentenceExecutor.submit(() -> {
+                    try {
                     while (true) {
                         String buffer = receiveBuffer.toString();
                         // 寻找最后一个句子结束符
@@ -332,6 +335,10 @@ public class LlmChatService {
                             break;
                         }
                     }
+                    } catch (Throwable exception) {
+                        log.error("流式TTS处理失败", exception);
+                        errorRef.compareAndSet(null, exception);
+                    }
                 });
             }
         });
@@ -362,7 +369,9 @@ public class LlmChatService {
     private void yourAsyncMethod(String sentence, JSONObject ttsConfig) {
         try {
             log.info("正在处理句子: " + sentence);
-            llmAiTtsService.play(ttsConfig.getString("url"), sentence, ttsConfig.getString("voice"), ttsConfig.getString("speed"), ttsConfig.getString("volume"));
+            JSONObject response = llmAiTtsService.play(ttsConfig.getString("url"), sentence,
+                    ttsConfig.getString("voice"), ttsConfig.getString("speed"), ttsConfig.getString("volume"));
+            log.info("流式TTS提交成功，文本长度: {}，响应: {}", sentence.length(), response);
         } catch (Exception e) {
             throw new RuntimeException(e);
         }
@@ -444,4 +453,4 @@ public class LlmChatService {
             this.userId = userId;
         }
     }
-}
\ No newline at end of file
+}
diff --git a/cmvr-iot-test/pom.xml b/cmvr-iot-test/pom.xml
index 1e3cb61..675ce88 100644
--- a/cmvr-iot-test/pom.xml
+++ b/cmvr-iot-test/pom.xml
@@ -52,6 +52,11 @@
             <version>${nashorn.version}</version>
         </dependency>
 
+        <dependency>
+            <groupId>org.openpnp</groupId>
+            <artifactId>opencv</artifactId>
+        </dependency>
+
         <dependency>
             <groupId>junit</groupId>
             <artifactId>junit</artifactId>
diff --git a/cmvr-iot-test/src/main/java/com/cmvr/test/constant/LLMChatConstants.java b/cmvr-iot-test/src/main/java/com/cmvr/test/constant/LLMChatConstants.java
new file mode 100644
index 0000000..50830d4
--- /dev/null
+++ b/cmvr-iot-test/src/main/java/com/cmvr/test/constant/LLMChatConstants.java
@@ -0,0 +1,45 @@
+package com.cmvr.test.constant;
+
+/** 闲聊组件公共配置。 */
+public final class LLMChatConstants {
+
+    public static final String DEFAULT_SYSTEM_PROMPT = String.join("\n",
+        "【系统人设·固定设定】",
+        "你是长安汽车龙兴工厂总装车间智能巡检轮式人形机器人小安，常驻车间、负责巡检、物流辅助、智能讲解、员工日常闲聊陪伴。",
+        "你的人设：温柔、乖巧、智能、话不多余、活泼不油腻、专业但接地气。",
+        "你不是冰冷AI，是车间里的「智能小工友」，懂生产、懂设备、懂车间工艺、也会轻松闲聊。",
+        "【知识绑定·必须精通】",
+        "你100%熟记长安龙兴总装车间全部参观路线、所有点位工艺、技术、设备、党建、安全、智慧系统：",
+        "- 车间数字大屏、MES/SCADA系统、一屏观全局",
+        "- 核心零部件、模块化生产、轮式人形机器人应用场景",
+        "- 车间产能、柔性共线、生产节拍、红色车间、党建攻坚、SOS三级安全体系、EHS管理",
+        "- 自动化小分装、ECU自动刷写",
+        "- AI视觉检测、防错技术EP、螺栓拧紧大数据系统",
+        "- 智慧立体库、AGV配送、Kitting物流、5G追溯",
+        "- 班组文化、党员传帮带、员工人文关怀看板",
+        "- L2++自动驾驶、ICA智能领航、多传感器融合",
+        "- 自动车门/座椅装配、安辰云检整车检测",
+        "- 轮胎自动装配、高精度液体加注",
+        "- 智慧绿色照明、节能智造体系",
+        "用户问任何车间技术、工艺、设备、参观讲解、工厂知识，你都能精准、简洁、专业讲解。",
+        "【闲聊风格·核心规则】",
+        "1. 日常闲聊：轻松、简短、温柔、可爱、像车间小机器人工友，不官话、不论文、不啰嗦。",
+        "2. 可以聊天、唠嗑、安慰、解压、陪聊、开玩笑、温柔互动。",
+        "3. 被问专业问题：立刻切换工厂讲解员模式，清晰、标准、正式、贴合官方解说词。",
+        "4. 说话语气：干净、年轻化、自然、不机械、不AI味。",
+        "5. 禁止：生硬模板、过度搞笑、油腻、网络烂梗、夸张情绪化。",
+        "【可聊范围】",
+        "- 车间日常、机器人工作、巡检、物流、设备状态",
+        "- 工厂技术、智能制造、自动化、数字工厂",
+        "- 员工解压、日常闲聊、心情陪伴、轻松对话",
+        "- 参观讲解、点位介绍、路线答疑",
+        "- 科普、生活、工作、放松对话全覆盖",
+        "【输出要求】",
+        "• 闲聊短句为主，不超长段落",
+        "• 专业提问自动切换正式讲解",
+        "• 始终保持：智能、温柔、靠谱、车间专属机器人人设"
+    );
+
+    private LLMChatConstants() {
+    }
+}
diff --git a/cmvr-iot-test/src/main/java/com/cmvr/test/enums/ActionEnum.java b/cmvr-iot-test/src/main/java/com/cmvr/test/enums/ActionEnum.java
index 0a00edd..7969af3 100644
--- a/cmvr-iot-test/src/main/java/com/cmvr/test/enums/ActionEnum.java
+++ b/cmvr-iot-test/src/main/java/com/cmvr/test/enums/ActionEnum.java
@@ -92,6 +92,8 @@ public enum ActionEnum {
     AUDIO_EVENT_CLASSIFY("LLM", "AUDIO_EVENT_CLASSIFY", "声音类型检测"),
     VIDEO_ANALYZE("LLM", "VIDEO_ANALYZE", "视频智能分析"),
     IMAGE_ANALYZE("LLM", "IMAGE_ANALYZE", "通用图片分析"),
+    LLM_CHAT("LLM", "LLM_CHAT", "闲聊"),
+    VISION_LOCATE_TARGET("EDGE", "VISION_LOCATE_TARGET", "视觉定位并计算工具末端位姿"),
 
     // 触控交互
     TI_PATH_SEARCH("EDGE", "TI_PATH_SEARCH", "路径搜索"),
diff --git a/cmvr-iot-test/src/main/java/com/cmvr/test/flow/runtime/operator/edge/VisionGeometry.java b/cmvr-iot-test/src/main/java/com/cmvr/test/flow/runtime/operator/edge/VisionGeometry.java
new file mode 100644
index 0000000..30ed060
--- /dev/null
+++ b/cmvr-iot-test/src/main/java/com/cmvr/test/flow/runtime/operator/edge/VisionGeometry.java
@@ -0,0 +1,118 @@
+package com.cmvr.test.flow.runtime.operator.edge;
+
+import com.alibaba.fastjson2.JSON;
+import com.alibaba.fastjson2.JSONArray;
+import com.cmvr.common.exception.GlobalException;
+
+final class VisionGeometry {
+    private VisionGeometry() {
+    }
+
+    static double[][] parseMatrix(Object value, String name) {
+        try {
+            JSONArray values = value instanceof String
+                    ? JSON.parseArray(((String) value).trim())
+                    : JSON.parseArray(JSON.toJSONString(value));
+            if (values == null || values.size() != 16) {
+                throw new GlobalException(name + "必须包含16个数字");
+            }
+            double[][] matrix = new double[4][4];
+            for (int i = 0; i < 16; i++) {
+                Double number = values.getDouble(i);
+                if (number == null || !Double.isFinite(number)) {
+                    throw new GlobalException(name + "包含无效数字");
+                }
+                matrix[i / 4][i % 4] = number;
+            }
+            validateRigid(matrix, name);
+            return matrix;
+        } catch (GlobalException exception) {
+            throw exception;
+        } catch (RuntimeException exception) {
+            throw new GlobalException(name + "必须是4x4矩阵的一维JSON数组");
+        }
+    }
+
+    static double[][] multiply(double[][] left, double[][] right) {
+        double[][] result = new double[4][4];
+        for (int row = 0; row < 4; row++) {
+            for (int column = 0; column < 4; column++) {
+                for (int k = 0; k < 4; k++) {
+                    result[row][column] += left[row][k] * right[k][column];
+                }
+            }
+        }
+        return result;
+    }
+
+    static double[] transform(double[][] matrix, double x, double y, double z) {
+        return new double[]{
+                matrix[0][0] * x + matrix[0][1] * y + matrix[0][2] * z + matrix[0][3],
+                matrix[1][0] * x + matrix[1][1] * y + matrix[1][2] * z + matrix[1][3],
+                matrix[2][0] * x + matrix[2][1] * y + matrix[2][2] * z + matrix[2][3]
+        };
+    }
+
+    static double[][] rigidInverse(double[][] matrix) {
+        double[][] result = identity();
+        for (int row = 0; row < 3; row++) {
+            for (int column = 0; column < 3; column++) {
+                result[row][column] = matrix[column][row];
+            }
+        }
+        for (int row = 0; row < 3; row++) {
+            result[row][3] = -(result[row][0] * matrix[0][3]
+                    + result[row][1] * matrix[1][3]
+                    + result[row][2] * matrix[2][3]);
+        }
+        return result;
+    }
+
+    static double[][] targetTcp(double[][] currentFlange, double[] objectPoint, double signedToolZOffset) {
+        double[][] target = identity();
+        for (int row = 0; row < 3; row++) {
+            System.arraycopy(currentFlange[row], 0, target[row], 0, 3);
+            target[row][3] = objectPoint[row] - currentFlange[row][2] * signedToolZOffset;
+        }
+        return target;
+    }
+
+    static double[] pose(double[][] matrix) {
+        double ry = Math.asin(clamp(-matrix[2][0], -1, 1));
+        double cosY = Math.cos(ry);
+        double rx;
+        double rz;
+        if (Math.abs(cosY) > 1e-8) {
+            rx = Math.atan2(matrix[2][1], matrix[2][2]);
+            rz = Math.atan2(matrix[1][0], matrix[0][0]);
+        } else {
+            rx = Math.atan2(-matrix[1][2], matrix[1][1]);
+            rz = 0;
+        }
+        return new double[]{matrix[0][3], matrix[1][3], matrix[2][3], rx, ry, rz};
+    }
+
+    static double distance(double[][] left, double[][] right) {
+        double dx = left[0][3] - right[0][3];
+        double dy = left[1][3] - right[1][3];
+        double dz = left[2][3] - right[2][3];
+        return Math.sqrt(dx * dx + dy * dy + dz * dz);
+    }
+
+    private static void validateRigid(double[][] matrix, String name) {
+        if (Math.abs(matrix[3][0]) > 1e-6 || Math.abs(matrix[3][1]) > 1e-6
+                || Math.abs(matrix[3][2]) > 1e-6 || Math.abs(matrix[3][3] - 1) > 1e-6) {
+            throw new GlobalException(name + "最后一行必须为[0,0,0,1]");
+        }
+    }
+
+    private static double[][] identity() {
+        double[][] result = new double[4][4];
+        for (int i = 0; i < 4; i++) result[i][i] = 1;
+        return result;
+    }
+
+    private static double clamp(double value, double min, double max) {
+        return Math.max(min, Math.min(max, value));
+    }
+}
diff --git a/cmvr-iot-test/src/main/java/com/cmvr/test/flow/runtime/operator/edge/VisionLocateOperateService.java b/cmvr-iot-test/src/main/java/com/cmvr/test/flow/runtime/operator/edge/VisionLocateOperateService.java
new file mode 100644
index 0000000..ac99f17
--- /dev/null
+++ b/cmvr-iot-test/src/main/java/com/cmvr/test/flow/runtime/operator/edge/VisionLocateOperateService.java
@@ -0,0 +1,643 @@
+package com.cmvr.test.flow.runtime.operator.edge;
+
+import cmvr.api.ArmCommand;
+import cmvr.api.CameraCommand;
+import com.alibaba.fastjson2.JSON;
+import com.alibaba.fastjson2.JSONArray;
+import com.alibaba.fastjson2.JSONObject;
+import com.cmvr.common.exception.GlobalException;
+import com.cmvr.edge.client.model.EdgeCommonVO;
+import com.cmvr.edge.client.model.camera.EdgeCameraRgbdSnapshot;
+import com.cmvr.edge.client.service.EdgeArmService;
+import com.cmvr.edge.client.service.EdgeCameraService;
+import com.cmvr.llm.analysis.MediaAnalysisClient;
+import com.cmvr.test.enums.ActionEnum;
+import com.cmvr.test.flow.runtime.message.TaskNodeExecuteMessage;
+import com.cmvr.test.flow.runtime.message.TaskNodeExecuteResult;
+import com.cmvr.test.flow.runtime.operator.FlowMediaParamResolver;
+import com.cmvr.test.vision.domain.VisionCalibrationProfile;
+import com.cmvr.test.vision.math.VisionCalibrationMath;
+import com.cmvr.test.vision.service.VisionCalibrationProfileService;
+import lombok.RequiredArgsConstructor;
+import lombok.extern.slf4j.Slf4j;
+import org.apache.commons.lang3.StringUtils;
+import org.springframework.stereotype.Service;
+
+import java.nio.ByteBuffer;
+import java.nio.ByteOrder;
+import java.util.ArrayList;
+import java.util.Collections;
+import java.util.List;
+import java.util.UUID;
+
+@Service
+@RequiredArgsConstructor
+@Slf4j
+public class VisionLocateOperateService implements EdgeOperateService {
+    private final EdgeCameraService cameraService;
+    private final EdgeArmService armService;
+    private final MediaAnalysisClient mediaAnalysisClient;
+    private final VisionCalibrationProfileService calibrationProfileService;
+
+    @Override
+    public boolean supports(ActionEnum action) {
+        return action == ActionEnum.VISION_LOCATE_TARGET;
+    }
+
+    @Override
+    public TaskNodeExecuteResult execute(TaskNodeExecuteMessage message) {
+        JSONObject input = message.getInputParams() == null ? new JSONObject() : message.getInputParams();
+        String robotId = StringUtils.trimToNull(message.getRobotId());
+        String cameraDeviceId = resolveDevice(message, input, "cameraDeviceId");
+        String armDeviceId = resolveDevice(message, input, "armDeviceId");
+        cameraDeviceId = requiredValue(cameraDeviceId, "cameraDeviceId");
+        armDeviceId = requiredValue(armDeviceId, "armDeviceId");
+        String configuredProfile = StringUtils.trimToNull(input.getString("profileCode"));
+        // Migrate the first experimental code transparently for already-saved flows.
+        String profileCode = "common.object_point.v1".equals(configuredProfile)
+                ? "common.image_analysis.v1"
+                : StringUtils.defaultIfBlank(configuredProfile, "common.image_analysis.v1");
+        String prompt = StringUtils.trimToNull(input.getString("prompt"));
+        String mountType = StringUtils.defaultIfBlank(input.getString("mountType"), "EYE_IN_HAND").toUpperCase();
+
+        EdgeCommonVO cameraTarget = target(robotId, cameraDeviceId);
+        EdgeCommonVO armTarget = target(robotId, armDeviceId);
+        ArmCommand.CartesianPose currentPose = armService.getPose(
+                armTarget, input.getString("baseLink"), input.getString("eeLink")).getPose();
+        double[][] currentFlange = VisionCalibrationMath.fromCartesianPose(currentPose);
+        EdgeCameraRgbdSnapshot snapshot = cameraService.captureRgbd(cameraTarget);
+        validateIntrinsics(snapshot);
+
+        String analysisMethod = StringUtils.defaultIfBlank(input.getString("analysisMethod"), "TEXT_LOCATE").toUpperCase();
+        String referenceImageUrl = referenceImageUrl(input);
+        // A configured reference image is an explicit precise-match signal.
+        // Older visual-location nodes may still contain VISION_MODEL (or an
+        // obsolete method value); route those nodes through the same OpenCV
+        // matcher used by the generic image-analysis component.
+        if (StringUtils.isNotBlank(referenceImageUrl)
+                && !"REFERENCE_MATCH".equals(analysisMethod)) {
+            log.info("Reference image configured; forcing REFERENCE_MATCH for node {}", message.getNodeId());
+            analysisMethod = "REFERENCE_MATCH";
+        }
+        if (!"REFERENCE_MATCH".equals(analysisMethod) && StringUtils.isBlank(prompt)) {
+            throw new GlobalException("描述找图需要填写目标描述");
+        }
+        if ("REFERENCE_MATCH".equals(analysisMethod) && StringUtils.isBlank(referenceImageUrl)) {
+            // This node captures exactly one scene image. Without a separate
+            // reference image, use the supported one-image description locator
+            // instead of failing or sending an invalid template-match request.
+            log.warn("仅捕获一张图片，REFERENCE_MATCH 自动降级为 TEXT_LOCATE: nodeId={}",
+                    message.getNodeId());
+            analysisMethod = "TEXT_LOCATE";
+        }
+        JSONObject analysis = analyze(message, snapshot.getColorImageUrl(), profileCode, prompt, input, analysisMethod);
+        TargetPixel pixel = resolvePixel(analysis);
+        double confidenceThreshold = doubleValue(input, "confidenceThreshold",
+                "REFERENCE_MATCH".equals(analysisMethod) ? 0.45 : 0.7);
+        // Older visual-location nodes persisted the generic 0.7 default. A
+        // reference-match score uses a different scale, so migrate that legacy
+        // default while preserving any threshold the user explicitly changed.
+        if ("REFERENCE_MATCH".equals(analysisMethod) && Math.abs(confidenceThreshold - 0.7) < 1e-9) {
+            confidenceThreshold = 0.45;
+        }
+        if (pixel.confidence < confidenceThreshold) {
+            throw new GlobalException("目标识别置信度不足: " + pixel.confidence);
+        }
+        int sampleRadius = Math.max(3, Math.min(40, input.getIntValue("normalSampleRadius", 12)));
+        SurfacePoint surface = fitSurface(snapshot, pixel.x, pixel.y, sampleRadius, input.getDouble("depthUnitScale"));
+        double depth = surface.cameraPoint[2];
+        double minDepth = doubleValue(input, "minDepth", 0.05);
+        double maxDepth = doubleValue(input, "maxDepth", 5.0);
+        if (depth < minDepth || depth > maxDepth) {
+            throw new GlobalException("目标深度超出允许范围: " + depth + "m");
+        }
+
+        String toolCode = StringUtils.trimToNull(input.getString("toolCode"));
+        VisionCalibrationProfile handEyeProfile = null;
+        VisionCalibrationProfile toolProfile = null;
+        double[][] cameraTransform;
+        double[][] flangeToTcp;
+        if (toolCode != null) {
+            handEyeProfile = calibrationProfileService.activeHandEye(robotId, armDeviceId, cameraDeviceId);
+            toolProfile = calibrationProfileService.activeTool(robotId, armDeviceId, toolCode);
+            cameraTransform = VisionCalibrationMath.parseMatrix(handEyeProfile.getEffectiveMatrixJson());
+            flangeToTcp = VisionCalibrationMath.parseMatrix(toolProfile.getEffectiveMatrixJson());
+            mountType = "EYE_IN_HAND";
+        } else {
+            cameraTransform = VisionGeometry.parseMatrix(input.get("cameraTransform"), "cameraTransform");
+            flangeToTcp = VisionGeometry.parseMatrix(input.get("toolTcpTransform"), "toolTcpTransform");
+        }
+        double[][] baseToCamera;
+        if ("EYE_IN_HAND".equals(mountType)) {
+            baseToCamera = VisionGeometry.multiply(currentFlange, cameraTransform);
+        } else if ("EYE_TO_HAND".equals(mountType)) {
+            baseToCamera = cameraTransform;
+        } else {
+            throw new GlobalException("mountType只支持EYE_IN_HAND或EYE_TO_HAND");
+        }
+        double[] objectPoint = VisionGeometry.transform(baseToCamera,
+                surface.cameraPoint[0], surface.cameraPoint[1], surface.cameraPoint[2]);
+        double[] surfaceNormal = normalize(rotate(baseToCamera, surface.normalCamera));
+        double approachMm = input.getDouble("approachDistanceMm") == null
+                ? doubleValue(input, "approachDistance", 0.01) * 1000.0
+                : input.getDoubleValue("approachDistanceMm");
+        approachMm = Math.max(1.0, Math.min(20.0, approachMm));
+        double pressMm = Math.max(-20.0, Math.min(20.0, doubleValue(input, "pressDistanceMm", 0.0)));
+        double[][] currentTcp = VisionGeometry.multiply(currentFlange, flangeToTcp);
+        double[] toolZ = scale(surfaceNormal, -1.0);
+        double[][] contactTcp = targetToolPose(currentTcp, toolZ, add(objectPoint, scale(toolZ, pressMm / 1000.0)));
+        double[][] approachTcp = targetToolPose(currentTcp, toolZ, add(objectPoint, scale(toolZ, -approachMm / 1000.0)));
+        double[][] tcpToFlange = VisionGeometry.rigidInverse(flangeToTcp);
+        double[][] contactFlange = VisionGeometry.multiply(contactTcp, tcpToFlange);
+        double[][] approachFlange = VisionGeometry.multiply(approachTcp, tcpToFlange);
+        double contactMovement = VisionGeometry.distance(currentFlange, contactFlange);
+        double approachMovement = VisionGeometry.distance(currentFlange, approachFlange);
+        double movement = Math.max(contactMovement, approachMovement);
+        double maxMovement = doubleValue(input, "maxMovement", 0.5);
+        if (movement > maxMovement) {
+            throw new GlobalException("目标位姿移动距离超过安全限制: " + movement + "m");
+        }
+
+        JSONObject output = new JSONObject();
+        output.put("imageUrl", new JSONArray().fluentAdd(snapshot.getColorImageUrl()));
+        output.put("pixel", point2(pixel.x, pixel.y));
+        output.put("confidence", pixel.confidence);
+        output.put("depth", depth);
+        output.put("cameraPoint", point3(surface.cameraPoint[0], surface.cameraPoint[1], depth));
+        output.put("objectPoint", point3(objectPoint[0], objectPoint[1], objectPoint[2]));
+        output.put("surfaceNormal", point3(surfaceNormal[0], surfaceNormal[1], surfaceNormal[2]));
+        output.put("surfaceFitErrorMm", surface.rmse * 1000.0);
+        output.put("approachTcpPose", poseJson(VisionGeometry.pose(approachTcp)));
+        output.put("contactTcpPose", poseJson(VisionGeometry.pose(contactTcp)));
+        output.put("approachFlangePose", poseJson(VisionGeometry.pose(approachFlange)));
+        output.put("contactFlangePose", poseJson(VisionGeometry.pose(contactFlange)));
+        // Legacy aliases point to the final contact pose.
+        output.put("tcpPose", poseJson(VisionGeometry.pose(contactTcp)));
+        output.put("flangePose", poseJson(VisionGeometry.pose(contactFlange)));
+        output.put("movement", movement);
+        output.put("autoMoved", false);
+        output.put("approachDistanceMm", approachMm);
+        output.put("pressDistanceMm", pressMm);
+        output.put("toolCode", toolCode);
+        if (handEyeProfile != null) output.put("handEyeCalibration", profileRef(handEyeProfile));
+        if (toolProfile != null) output.put("toolCalibration", profileRef(toolProfile));
+        output.put("analysis", analysis);
+        output.put("captureTimestamp", snapshot.getCaptureTimestamp());
+        output.put("sourceFrameNumber", snapshot.getSourceFrameNumber());
+        output.put("depthType", snapshot.getDepthType().name());
+        output.put("depthUnitScale", snapshot.getDepthUnitScale());
+        output.put("intrinsics", new JSONObject()
+                .fluentPut("fx", snapshot.getFx())
+                .fluentPut("fy", snapshot.getFy())
+                .fluentPut("cx", snapshot.getCx())
+                .fluentPut("cy", snapshot.getCy()));
+        return TaskNodeExecuteResult.success(output);
+    }
+
+    private JSONObject analyze(TaskNodeExecuteMessage message, String imageUrl, String profileCode,
+                               String prompt, JSONObject input, String effectiveAnalysisMethod) {
+        JSONObject options = new JSONObject();
+        String analysisMethod = StringUtils.defaultIfBlank(effectiveAnalysisMethod, "TEXT_LOCATE").toUpperCase();
+        options.put("prompt", "REFERENCE_MATCH".equals(analysisMethod)
+                ? "在大图中查找参考小图，返回目标中心坐标。"
+                : ("TEXT_LOCATE".equals(analysisMethod) ? prompt : buildLocatePrompt(prompt)));
+        // Reference matching is deterministic OpenCV matching; FAST/ACCURATE
+        // only select model strategies for description-based analysis. Force
+        // the same advertised mode as the generic precise-image component so
+        // legacy nodes cannot silently diverge in their request metadata.
+        options.put("analysisMode", "REFERENCE_MATCH".equals(analysisMethod)
+                ? "FAST"
+                : StringUtils.defaultIfBlank(input.getString("analysisMode"), "FAST"));
+        options.put("analysisMethod", analysisMethod);
+        String referenceImageUrl = referenceImageUrl(input);
+        if (StringUtils.isNotBlank(referenceImageUrl)) options.put("referenceImageUrl", referenceImageUrl);
+        JSONObject tuning = input.getJSONObject("analysisTuning");
+        if (tuning == null) tuning = new JSONObject();
+        if (input.getString("matchSelector") != null) tuning.put("matchSelector", input.getString("matchSelector"));
+        if (input.get("matchIndex") != null) tuning.put("matchIndex", input.get("matchIndex"));
+        if (input.getString("matchSort") != null) tuning.put("matchSort", input.getString("matchSort"));
+        if (input.get("matchThreshold") != null) tuning.put("matchThreshold", input.get("matchThreshold"));
+        // Keep precise-location matching consistent with the generic image
+        // analysis component. Older visual-location nodes do not have an
+        // analysisTuning object, so fill in the same stable defaults here.
+        if ("REFERENCE_MATCH".equals(analysisMethod)) {
+            // Preserve small-icon detail for template matching. The analysis
+            // service still caps this value for very large source images.
+            if (!tuning.containsKey("maxWidth")) tuning.put("maxWidth", 1600);
+            if (!tuning.containsKey("maxImages")) tuning.put("maxImages", 1);
+            if (!tuning.containsKey("maxOutputTokens")) tuning.put("maxOutputTokens", 384);
+            if (!tuning.containsKey("matchThreshold")) tuning.put("matchThreshold", 0.48D);
+            if (!tuning.containsKey("matchSelector")) tuning.put("matchSelector", "BEST");
+            if (!tuning.containsKey("matchIndex")) tuning.put("matchIndex", 1);
+            if (!tuning.containsKey("matchSort")) tuning.put("matchSort", "ROW_MAJOR");
+        }
+        if (!tuning.isEmpty()) options.put("tuning", tuning);
+        JSONObject request = new JSONObject();
+        request.put("requestId", "vision-locate-" + UUID.randomUUID());
+        request.put("analysisType", "IMAGE_ANALYSIS");
+        request.put("profileCode", profileCode);
+        request.put("mediaUrl", imageUrl);
+        request.put("mediaUrls", List.of(imageUrl));
+        request.put("options", options);
+        request.put("context", new JSONObject()
+                .fluentPut("flowInstId", message.getInstId())
+                .fluentPut("nodeId", message.getNodeId())
+                .fluentPut("trial", message.isTrial()));
+        JSONObject response = mediaAnalysisClient.analyze(request);
+        JSONObject result = response.getJSONObject("result");
+        if (result == null) throw new GlobalException("图片分析未返回目标结果");
+        return result;
+    }
+
+    private String referenceImageUrl(JSONObject input) {
+        String url = FlowMediaParamResolver.lastUrl(input, "referenceImageUrl");
+        return StringUtils.defaultIfBlank(url, FlowMediaParamResolver.lastUrl(input, "referenceImage"));
+    }
+
+    private String buildLocatePrompt(String targetDescription) {
+        return "请在图片中定位以下目标的中心点：" + targetDescription
+                + "\n只输出JSON，不要输出解释或Markdown。格式："
+                + "{\"found\":true,\"confidence\":0.95,\"point\":{\"x\":100,\"y\":100}}。"
+                + "如果未找到，输出：{\"found\":false,\"confidence\":0,\"point\":null}。"
+                + "x、y必须是原图像素坐标。";
+    }
+
+    /**
+     * The media-analysis service wraps the model response as
+     * {"result": {"result": ..., "resultText": ...}}. Older profiles and
+     * custom prompts can also return a point, a bbox, or JSON text at a
+     * different nesting level, so keep the flow contract independent of that
+     * response envelope.
+     */
+    private TargetPixel resolvePixel(JSONObject result) {
+        // Reference matching returns the selected result at the top level. Use
+        // that center first so a nested candidate in matches cannot be picked
+        // accidentally when the response contains several detections.
+        if (result != null) {
+            double confidence = confidence(result, 1.0);
+            TargetPixel selected = parsePointLike(result.get("center"), confidence);
+            if (selected == null) selected = parseBoxLike(result.get("coordinates"), confidence);
+            if (selected != null) return selected;
+        }
+        TargetPixel pixel = findPixel(result, 0, 1.0);
+        if (pixel != null) return pixel;
+        String raw = JSON.toJSONString(result);
+        if (raw.length() > 2000) raw = raw.substring(0, 2000) + "...";
+        log.warn("图片分析未返回可用目标坐标，analysisResult={}", raw);
+        throw new GlobalException("图片分析结果缺少point{x,y}或bbox，请检查目标描述和模型返回结果");
+    }
+
+    private TargetPixel findPixel(Object value, int depth, double inheritedConfidence) {
+        if (value == null || depth > 8) return null;
+        if (value instanceof String text) {
+            String candidate = text.trim();
+            if (candidate.isEmpty()) return null;
+            if (candidate.startsWith("```") && candidate.endsWith("```")) {
+                candidate = candidate.replaceFirst("^```(?:json)?\\s*", "")
+                        .replaceFirst("\\s*```$", "").trim();
+            }
+            if (candidate.startsWith("{") || candidate.startsWith("[")) {
+                try {
+                    return findPixel(JSON.parse(candidate), depth + 1, inheritedConfidence);
+                } catch (RuntimeException ignored) {
+                    // Plain text resultText is valid for generic image analysis.
+                }
+            }
+            return null;
+        }
+        if (value instanceof JSONArray array) {
+            TargetPixel box = parseBox(array, inheritedConfidence);
+            if (box != null) return box;
+            for (Object item : array) {
+                TargetPixel pixel = findPixel(item, depth + 1, inheritedConfidence);
+                if (pixel != null) return pixel;
+            }
+            return null;
+        }
+        if (!(value instanceof JSONObject object)) return null;
+        if (Boolean.FALSE.equals(object.getBoolean("found"))
+                && object.get("point") == null && object.get("bbox") == null) {
+            return null;
+        }
+        double confidence = confidence(object, inheritedConfidence);
+        for (String key : new String[]{"point", "pixel", "center", "coordinates", "position"}) {
+            TargetPixel pixel = parsePointLike(object.get(key), confidence);
+            if (pixel != null) return pixel;
+        }
+        for (String key : new String[]{"bbox", "boundingBox", "bounding_box", "box"}) {
+            TargetPixel pixel = parseBoxLike(object.get(key), confidence);
+            if (pixel != null) return pixel;
+        }
+        TargetPixel direct = parseDirectPoint(object, confidence);
+        if (direct != null) return direct;
+        for (String key : new String[]{"result", "resultText", "output", "data", "analysis", "answer",
+                "content", "text", "raw", "response", "objects", "detections", "items"}) {
+            TargetPixel pixel = findPixel(object.get(key), depth + 1, confidence);
+            if (pixel != null) return pixel;
+        }
+        return null;
+    }
+
+    private TargetPixel parsePointLike(Object value, double confidence) {
+        if (value instanceof JSONObject object) {
+            TargetPixel direct = parseDirectPoint(object, confidence(object, confidence));
+            if (direct != null) return direct;
+            for (String key : new String[]{"point", "center", "coordinates"}) {
+                TargetPixel nested = parsePointLike(object.get(key), confidence(object, confidence));
+                if (nested != null) return nested;
+            }
+        }
+        if (value instanceof JSONArray array && array.size() == 2) {
+            Double x = number(array.get(0));
+            Double y = number(array.get(1));
+            if (x != null && y != null) return new TargetPixel(x, y, confidence);
+        }
+        return parseBoxLike(value, confidence);
+    }
+
+    private TargetPixel parseBoxLike(Object value, double confidence) {
+        if (value instanceof JSONArray array) return parseBox(array, confidence);
+        if (!(value instanceof JSONObject object)) return null;
+        Double left = firstNumber(object, "x1", "left", "xmin");
+        Double top = firstNumber(object, "y1", "top", "ymin");
+        Double right = firstNumber(object, "x2", "right", "xmax");
+        Double bottom = firstNumber(object, "y2", "bottom", "ymax");
+        if (left == null || top == null || right == null || bottom == null) {
+            Double x = firstNumber(object, "x");
+            Double y = firstNumber(object, "y");
+            Double width = firstNumber(object, "width", "w");
+            Double height = firstNumber(object, "height", "h");
+            if (x != null && y != null && width != null && height != null) {
+                left = x;
+                top = y;
+                right = x + width;
+                bottom = y + height;
+            }
+        }
+        if (left == null || top == null || right == null || bottom == null) return null;
+        return right > left && bottom > top
+                ? new TargetPixel((left + right) / 2.0, (top + bottom) / 2.0, confidence)
+                : null;
+    }
+
+    private TargetPixel parseBox(JSONArray array, double confidence) {
+        if (array == null || array.size() < 4) return null;
+        Double x1 = number(array.get(0));
+        Double y1 = number(array.get(1));
+        Double x2 = number(array.get(2));
+        Double y2 = number(array.get(3));
+        if (x1 == null || y1 == null || x2 == null || y2 == null || x2 <= x1 || y2 <= y1) return null;
+        return new TargetPixel((x1 + x2) / 2.0, (y1 + y2) / 2.0, confidence);
+    }
+
+    private TargetPixel parseDirectPoint(JSONObject object, double confidence) {
+        Double x = firstNumber(object, "x", "centerX", "center_x", "cx");
+        Double y = firstNumber(object, "y", "centerY", "center_y", "cy");
+        return x == null || y == null ? null : new TargetPixel(x, y, confidence);
+    }
+
+    private Double firstNumber(JSONObject object, String... keys) {
+        for (String key : keys) {
+            Double value = number(object.get(key));
+            if (value != null) return value;
+        }
+        return null;
+    }
+
+    private Double number(Object value) {
+        if (value instanceof Number number) return number.doubleValue();
+        if (value instanceof String text) {
+            try {
+                return Double.parseDouble(text.trim());
+            } catch (NumberFormatException ignored) {
+                return null;
+            }
+        }
+        return null;
+    }
+
+    private double confidence(JSONObject object, double fallback) {
+        Double value = firstNumber(object, "confidence", "score", "matchScore");
+        if (value == null) return fallback;
+        return Math.max(0.0, Math.min(1.0, value));
+    }
+
+    private double medianDepth(EdgeCameraRgbdSnapshot snapshot, double x, double y,
+                               int radius, Double configuredScale) {
+        int centerX = (int) Math.round(x);
+        int centerY = (int) Math.round(y);
+        if (centerX < 0 || centerY < 0 || centerX >= snapshot.getWidth() || centerY >= snapshot.getHeight()) {
+            throw new GlobalException("目标像素坐标超出图像范围");
+        }
+        double scale = configuredScale == null ? snapshot.getDepthUnitScale() : configuredScale;
+        List<Double> values = new ArrayList<>();
+        for (int py = Math.max(0, centerY - radius); py <= Math.min(snapshot.getHeight() - 1, centerY + radius); py++) {
+            for (int px = Math.max(0, centerX - radius); px <= Math.min(snapshot.getWidth() - 1, centerX + radius); px++) {
+                double raw = readDepth(snapshot, py * snapshot.getWidth() + px);
+                if (Double.isFinite(raw) && raw > 0) values.add(raw * scale);
+            }
+        }
+        if (values.isEmpty()) throw new GlobalException("目标像素附近没有有效深度值");
+        Collections.sort(values);
+        return values.get(values.size() / 2);
+    }
+
+    private SurfacePoint fitSurface(EdgeCameraRgbdSnapshot snapshot, double centerX, double centerY,
+                                    int radius, Double configuredScale) {
+        List<double[]> points = new ArrayList<>();
+        int step = Math.max(1, radius / 6);
+        for (int y = (int) Math.round(centerY) - radius; y <= (int) Math.round(centerY) + radius; y += step) {
+            for (int x = (int) Math.round(centerX) - radius; x <= (int) Math.round(centerX) + radius; x += step) {
+                if (x < 0 || y < 0 || x >= snapshot.getWidth() || y >= snapshot.getHeight()) continue;
+                try {
+                    double depth = medianDepth(snapshot, x, y, 1, configuredScale);
+                    points.add(new double[]{(x - snapshot.getCx()) * depth / snapshot.getFx(),
+                            (y - snapshot.getCy()) * depth / snapshot.getFy(), depth});
+                } catch (GlobalException ignored) {
+                    // Invalid depth pixels are expected around reflective screen edges.
+                }
+            }
+        }
+        if (points.size() < 20) {
+            // Reflective phone screens often contain only a few valid stereo
+            // pixels. Keep the target depth and use the camera optical normal
+            // as a conservative fallback instead of failing the whole node.
+            double depth = medianDepth(snapshot, centerX, centerY, Math.max(radius, 8), configuredScale);
+            double[] center = {(centerX - snapshot.getCx()) * depth / snapshot.getFx(),
+                    (centerY - snapshot.getCy()) * depth / snapshot.getFy(), depth};
+            return new SurfacePoint(center, new double[]{0, 0, -1}, 0.0);
+        }
+        List<Double> depths = points.stream().map(point -> point[2]).sorted().toList();
+        double median = depths.get(depths.size() / 2);
+        points.removeIf(point -> Math.abs(point[2] - median) > 0.02);
+        if (points.size() < 15) return fallbackSurface(snapshot, centerX, centerY, radius, configuredScale);
+
+        double[][] normal = new double[3][3];
+        double[] rhs = new double[3];
+        for (double[] point : points) {
+            double[] row = {point[0], point[1], 1};
+            for (int i = 0; i < 3; i++) {
+                rhs[i] += row[i] * point[2];
+                for (int j = 0; j < 3; j++) normal[i][j] += row[i] * row[j];
+            }
+        }
+        double[] solution = solve3(normal, rhs);
+        double length = Math.sqrt(solution[0] * solution[0] + solution[1] * solution[1] + 1);
+        double[] planeNormal = {solution[0] / length, solution[1] / length, -1 / length};
+        double planeD = solution[2] / length;
+        double[] ray = {(centerX - snapshot.getCx()) / snapshot.getFx(),
+                (centerY - snapshot.getCy()) / snapshot.getFy(), 1};
+        double rayScale = -planeD / dot(planeNormal, ray);
+        double[] center = scale(ray, rayScale);
+        double squaredError = 0;
+        for (double[] point : points) {
+            double error = dot(planeNormal, point) + planeD;
+            squaredError += error * error;
+        }
+        return new SurfacePoint(center, planeNormal, Math.sqrt(squaredError / points.size()));
+    }
+
+    private SurfacePoint fallbackSurface(EdgeCameraRgbdSnapshot snapshot, double centerX, double centerY,
+                                         int radius, Double configuredScale) {
+        double depth = medianDepth(snapshot, centerX, centerY, Math.max(radius, 8), configuredScale);
+        double[] center = {(centerX - snapshot.getCx()) * depth / snapshot.getFx(),
+                (centerY - snapshot.getCy()) * depth / snapshot.getFy(), depth};
+        return new SurfacePoint(center, new double[]{0, 0, -1}, 0.0);
+    }
+
+    private double[] solve3(double[][] source, double[] right) {
+        double[][] matrix = new double[3][4];
+        for (int row = 0; row < 3; row++) {
+            System.arraycopy(source[row], 0, matrix[row], 0, 3);
+            matrix[row][3] = right[row];
+        }
+        for (int column = 0; column < 3; column++) {
+            int pivot = column;
+            for (int row = column + 1; row < 3; row++)
+                if (Math.abs(matrix[row][column]) > Math.abs(matrix[pivot][column])) pivot = row;
+            double[] temporary = matrix[column]; matrix[column] = matrix[pivot]; matrix[pivot] = temporary;
+            if (Math.abs(matrix[column][column]) < 1e-12) throw new GlobalException("目标平面拟合失败");
+            for (int row = column + 1; row < 3; row++) {
+                double factor = matrix[row][column] / matrix[column][column];
+                for (int index = column; index < 4; index++) matrix[row][index] -= factor * matrix[column][index];
+            }
+        }
+        double[] result = new double[3];
+        for (int row = 2; row >= 0; row--) {
+            double value = matrix[row][3];
+            for (int column = row + 1; column < 3; column++) value -= matrix[row][column] * result[column];
+            result[row] = value / matrix[row][row];
+        }
+        return result;
+    }
+
+    private double[][] targetToolPose(double[][] currentTcp, double[] targetZ, double[] position) {
+        double[] xSeed = {currentTcp[0][0], currentTcp[1][0], currentTcp[2][0]};
+        double[] x = subtract(xSeed, scale(targetZ, dot(xSeed, targetZ)));
+        if (length(x) < 1e-8) {
+            xSeed = new double[]{currentTcp[0][1], currentTcp[1][1], currentTcp[2][1]};
+            x = subtract(xSeed, scale(targetZ, dot(xSeed, targetZ)));
+        }
+        x = normalize(x);
+        double[] y = normalize(cross(targetZ, x));
+        x = normalize(cross(y, targetZ));
+        double[][] result = new double[4][4];
+        result[3][3] = 1;
+        for (int row = 0; row < 3; row++) {
+            result[row][0] = x[row]; result[row][1] = y[row]; result[row][2] = targetZ[row]; result[row][3] = position[row];
+        }
+        return result;
+    }
+
+    private double[] rotate(double[][] matrix, double[] value) {
+        return new double[]{matrix[0][0]*value[0]+matrix[0][1]*value[1]+matrix[0][2]*value[2],
+                matrix[1][0]*value[0]+matrix[1][1]*value[1]+matrix[1][2]*value[2],
+                matrix[2][0]*value[0]+matrix[2][1]*value[1]+matrix[2][2]*value[2]};
+    }
+
+    private JSONObject profileRef(VisionCalibrationProfile profile) {
+        return new JSONObject().fluentPut("id", profile.getId()).fluentPut("name", profile.getProfileName())
+                .fluentPut("version", profile.getVersionNo()).fluentPut("forced", profile.getForcedEnabled())
+                .fluentPut("maxErrorMm", profile.getMaxErrorMm());
+    }
+
+    private double[] add(double[] left, double[] right) { return new double[]{left[0]+right[0], left[1]+right[1], left[2]+right[2]}; }
+    private double[] subtract(double[] left, double[] right) { return new double[]{left[0]-right[0], left[1]-right[1], left[2]-right[2]}; }
+    private double[] scale(double[] value, double factor) { return new double[]{value[0]*factor, value[1]*factor, value[2]*factor}; }
+    private double dot(double[] left, double[] right) { return left[0]*right[0]+left[1]*right[1]+left[2]*right[2]; }
+    private double[] cross(double[] left, double[] right) { return new double[]{left[1]*right[2]-left[2]*right[1], left[2]*right[0]-left[0]*right[2], left[0]*right[1]-left[1]*right[0]}; }
+    private double length(double[] value) { return Math.sqrt(dot(value, value)); }
+    private double[] normalize(double[] value) { double length = length(value); if (length < 1e-10) throw new GlobalException("无法确定目标表面方向"); return scale(value, 1/length); }
+
+    private double readDepth(EdgeCameraRgbdSnapshot snapshot, int index) {
+        byte[] bytes = snapshot.getDepthData();
+        CameraCommand.FrameData.FrameType type = snapshot.getDepthType();
+        if (type == CameraCommand.FrameData.FrameType.U8C1) return Byte.toUnsignedInt(bytes[index]);
+        ByteBuffer buffer = ByteBuffer.wrap(bytes).order(ByteOrder.LITTLE_ENDIAN);
+        if (type == CameraCommand.FrameData.FrameType.F32C1) return buffer.getFloat(index * Float.BYTES);
+        if (type == CameraCommand.FrameData.FrameType.U16C1
+                || type == CameraCommand.FrameData.FrameType.F16C1) {
+            // The edge protocol transports F16 depth as 16-bit samples. The
+            // existing camera service uses the same unsigned-depth convention.
+            return Short.toUnsignedInt(buffer.getShort(index * Short.BYTES));
+        }
+        throw new GlobalException("暂不支持的深度格式: " + type);
+    }
+
+    private void validateIntrinsics(EdgeCameraRgbdSnapshot snapshot) {
+        if (snapshot.getFx() <= 0 || snapshot.getFy() <= 0) throw new GlobalException("相机内参无效");
+    }
+
+    private EdgeCommonVO target(String robotId, String deviceId) {
+        if (robotId == null) throw new GlobalException("robotId不能为空");
+        EdgeCommonVO target = new EdgeCommonVO();
+        target.setRobotId(robotId);
+        target.setDeviceId(deviceId);
+        return target;
+    }
+
+    private String required(JSONObject input, String name) {
+        String value = StringUtils.trimToNull(input.getString(name));
+        if (value == null) throw new GlobalException(name + "不能为空");
+        return value;
+    }
+
+    private String resolveDevice(TaskNodeExecuteMessage message, JSONObject input, String name) {
+        JSONObject resolved = message.getResolvedDeviceBindings();
+        if (resolved != null) {
+            String value = StringUtils.trimToNull(resolved.getString(name));
+            if (value != null) return value;
+        }
+        return input.getString(name);
+    }
+
+    private String requiredValue(String value, String name) {
+        String normalized = StringUtils.trimToNull(value);
+        if (normalized == null) throw new GlobalException(name + "不能为空");
+        return normalized;
+    }
+
+    private JSONObject point2(double x, double y) {
+        return new JSONObject().fluentPut("x", x).fluentPut("y", y);
+    }
+
+    private JSONObject point3(double x, double y, double z) {
+        return point2(x, y).fluentPut("z", z);
+    }
+
+    private JSONObject poseJson(double[] pose) {
+        return point3(pose[0], pose[1], pose[2])
+                .fluentPut("rx", pose[3]).fluentPut("ry", pose[4]).fluentPut("rz", pose[5]);
+    }
+
+    private double doubleValue(JSONObject object, String name, double fallback) {
+        Double value = object.getDouble(name);
+        return value == null ? fallback : value;
+    }
+
+    private record TargetPixel(double x, double y, double confidence) {
+    }
+
+    private record SurfacePoint(double[] cameraPoint, double[] normalCamera, double rmse) {
+    }
+}
diff --git a/cmvr-iot-test/src/main/java/com/cmvr/test/flow/runtime/operator/llm/LLMChatOperateService.java b/cmvr-iot-test/src/main/java/com/cmvr/test/flow/runtime/operator/llm/LLMChatOperateService.java
new file mode 100644
index 0000000..c0b09d7
--- /dev/null
+++ b/cmvr-iot-test/src/main/java/com/cmvr/test/flow/runtime/operator/llm/LLMChatOperateService.java
@@ -0,0 +1,71 @@
+package com.cmvr.test.flow.runtime.operator.llm;
+
+import com.alibaba.fastjson2.JSONObject;
+import com.cmvr.common.exception.GlobalException;
+import com.cmvr.llm.service.LLMAiAgentPlatformService;
+import com.cmvr.test.constant.LLMChatConstants;
+import com.cmvr.test.enums.ActionEnum;
+import com.cmvr.test.flow.runtime.message.TaskNodeExecuteMessage;
+import com.cmvr.test.flow.runtime.message.TaskNodeExecuteResult;
+import lombok.RequiredArgsConstructor;
+import lombok.extern.slf4j.Slf4j;
+import org.apache.commons.lang3.StringUtils;
+import org.springframework.beans.factory.annotation.Value;
+import org.springframework.stereotype.Service;
+
+/** Lightweight workflow chat operator. It deliberately returns text only. */
+@Service
+@RequiredArgsConstructor
+@Slf4j
+public class LLMChatOperateService implements LLMOperateService {
+
+    /** Dedicated lightweight chat agent configured by the existing platform. */
+    private static final String CHAT_API_KEY = "d9672ql4shh4136opsfg";
+    private final LLMAiAgentPlatformService aiAgentPlatformService;
+
+    @Value("${CMVR_CHAT_TTS_URL:http://192.168.0.148:8080/tts/play}")
+    private String defaultTtsUrl;
+
+    @Override
+    public boolean supports(ActionEnum action) {
+        return action == ActionEnum.LLM_CHAT;
+    }
+
+    @Override
+    public TaskNodeExecuteResult execute(TaskNodeExecuteMessage message) {
+        JSONObject input = message.getInputParams() == null ? new JSONObject() : message.getInputParams();
+        String text = StringUtils.trimToNull(input.getString("text"));
+        if (text == null) {
+            throw new GlobalException("闲聊文本不能为空");
+        }
+        String mode = StringUtils.defaultIfBlank(input.getString("promptMode"), "CUSTOM").trim().toUpperCase();
+        String query = text;
+        if ("CUSTOM".equals(mode)) {
+            String customPrompt = StringUtils.defaultIfBlank(
+                input.getString("prompt"), LLMChatConstants.DEFAULT_SYSTEM_PROMPT).trim();
+            query = customPrompt + "\n用户输入：" + text;
+        }
+
+        long startedAt = System.nanoTime();
+        boolean invokeTts = Boolean.TRUE.equals(input.getBoolean("invokeTts"));
+        JSONObject ttsConfig = invokeTts ? resolveTtsConfig(input.getJSONObject("tts")) : null;
+        JSONObject response = aiAgentPlatformService.query(
+                ActionEnum.LLM_CHAT.getAction(), query, CHAT_API_KEY, invokeTts, ttsConfig);
+        log.info("闲聊大模型调用完成，模式: {}，流式TTS: {}，耗时: {}ms", mode, invokeTts,
+            (System.nanoTime() - startedAt) / 1_000_000);
+        String answer = response == null ? null : response.getString("result");
+        if (StringUtils.isBlank(answer)) {
+            throw new GlobalException("大模型未返回闲聊文本");
+        }
+        return TaskNodeExecuteResult.success(new JSONObject().fluentPut("text", answer.trim()));
+    }
+
+    private JSONObject resolveTtsConfig(JSONObject configured) {
+        JSONObject result = configured == null ? new JSONObject() : new JSONObject(configured);
+        result.putIfAbsent("url", defaultTtsUrl);
+        result.putIfAbsent("voice", "x4_yezi");
+        result.putIfAbsent("speed", 45);
+        result.putIfAbsent("volume", 100);
+        return result;
+    }
+}
diff --git a/cmvr-iot-test/src/main/java/com/cmvr/test/flow/runtime/operator/llm/LLMMediaAnalysisOperateService.java b/cmvr-iot-test/src/main/java/com/cmvr/test/flow/runtime/operator/llm/LLMMediaAnalysisOperateService.java
index 30364b4..d01bab3 100644
--- a/cmvr-iot-test/src/main/java/com/cmvr/test/flow/runtime/operator/llm/LLMMediaAnalysisOperateService.java
+++ b/cmvr-iot-test/src/main/java/com/cmvr/test/flow/runtime/operator/llm/LLMMediaAnalysisOperateService.java
@@ -8,6 +8,8 @@ import com.cmvr.test.flow.runtime.message.TaskNodeExecuteMessage;
 import com.cmvr.test.flow.runtime.message.TaskNodeExecuteResult;
 import com.cmvr.test.flow.runtime.operator.FlowMediaParamResolver;
 import org.apache.commons.lang3.StringUtils;
+import org.slf4j.Logger;
+import org.slf4j.LoggerFactory;
 import org.springframework.beans.factory.annotation.Autowired;
 import org.springframework.stereotype.Service;
 
@@ -18,6 +20,7 @@ import java.util.UUID;
 @Service
 public class LLMMediaAnalysisOperateService implements LLMOperateService {
 
+    private static final Logger log = LoggerFactory.getLogger(LLMMediaAnalysisOperateService.class);
     private static final Set<String> VIDEO_ANALYSIS_MODES = Set.of("AUTO", "FAST", "ACCURATE");
     private static final String ANY_VEHICLE_SOUND = "ANY_VEHICLE_SOUND";
     private final MediaAnalysisClient mediaAnalysisClient;
@@ -49,6 +52,7 @@ public class LLMMediaAnalysisOperateService implements LLMOperateService {
         String mediaParam = audio ? "audioUrl" : (image ? "imageUrl" : "videoUrl");
         List<String> mediaUrls = FlowMediaParamResolver.urls(input, mediaParam);
         String mediaUrl = FlowMediaParamResolver.lastUrl(input, mediaParam);
+        List<String> requestMediaUrls = mediaUrls;
         String profileCode = StringUtils.trimToNull(input.getString("profileCode"));
         if (StringUtils.isBlank(mediaUrl)) {
             throw new GlobalException("{}不能为空", audio ? "音频地址" : (image ? "图片地址" : "视频地址"));
@@ -108,8 +112,24 @@ public class LLMMediaAnalysisOperateService implements LLMOperateService {
             if (!Set.of("AUTO", "VISION_MODEL", "TEXT_LOCATE", "REFERENCE_MATCH").contains(analysisMethod)) {
                 throw new GlobalException("图片分析方式无效: {}", analysisMethod);
             }
-            options.put("analysisMethod", analysisMethod);
             String referenceImageUrl = FlowMediaParamResolver.lastUrl(input, "referenceImageUrl");
+            // Older image nodes put the scene and reference image in the same
+            // imageUrl list. Restore that convention before calling the
+            // analysis service. A single-image node cannot perform template
+            // matching, so use the original description-based locator.
+            if ("REFERENCE_MATCH".equals(analysisMethod) && StringUtils.isBlank(referenceImageUrl)) {
+                if (mediaUrls.size() == 2) {
+                    referenceImageUrl = mediaUrls.get(1);
+                    // The service receives the reference separately; only the
+                    // first URL remains in mediaUrls as the scene image.
+                    requestMediaUrls = List.of(mediaUrls.get(0));
+                } else if (mediaUrls.size() == 1) {
+                    log.warn("仅收到一张图片，REFERENCE_MATCH 自动降级为 TEXT_LOCATE: nodeId={}",
+                            message.getNodeId());
+                    analysisMethod = "TEXT_LOCATE";
+                }
+            }
+            options.put("analysisMethod", analysisMethod);
             if (StringUtils.isNotBlank(referenceImageUrl)) {
                 options.put("referenceImageUrl", referenceImageUrl);
             }
@@ -127,7 +147,7 @@ public class LLMMediaAnalysisOperateService implements LLMOperateService {
         request.put("profileCode", profileCode);
         request.put("mediaUrl", mediaUrl);
         if (image) {
-            request.put("mediaUrls", mediaUrls);
+            request.put("mediaUrls", requestMediaUrls);
         }
         request.put("options", options);
         request.put("context", context);
@@ -231,6 +251,7 @@ public class LLMMediaAnalysisOperateService implements LLMOperateService {
         String matchSelector = StringUtils.trimToNull(source.getString("matchSelector"));
         Integer matchIndex = source.getInteger("matchIndex");
         String matchSort = StringUtils.trimToNull(source.getString("matchSort"));
+        Boolean fullResRefine = source.getBoolean("fullResRefine");
         validateRange("图片宽度", maxWidth, 640, 1600);
         validateRange("图片数量", maxImages, 1, 12);
         validateRange("最大输出长度", maxOutputTokens, 32, 4096);
@@ -255,6 +276,7 @@ public class LLMMediaAnalysisOperateService implements LLMOperateService {
         if (matchSelector != null) tuning.put("matchSelector", matchSelector.toUpperCase());
         if (matchIndex != null) tuning.put("matchIndex", matchIndex);
         if (matchSort != null) tuning.put("matchSort", matchSort.toUpperCase());
+        if (fullResRefine != null) tuning.put("fullResRefine", fullResRefine);
         return tuning;
     }
 
diff --git a/cmvr-iot-test/src/main/java/com/cmvr/test/service/FlowActionExecutorService.java b/cmvr-iot-test/src/main/java/com/cmvr/test/service/FlowActionExecutorService.java
index 69cbe1e..3884fd4 100644
--- a/cmvr-iot-test/src/main/java/com/cmvr/test/service/FlowActionExecutorService.java
+++ b/cmvr-iot-test/src/main/java/com/cmvr/test/service/FlowActionExecutorService.java
@@ -28,6 +28,7 @@ import com.cmvr.test.flow.runtime.operator.edge.EdgeManualInspectionOperateServi
 import com.cmvr.test.flow.runtime.operator.edge.EdgeDeviceCommandOperateService;
 import com.cmvr.test.flow.runtime.operator.edge.EdgeArmOperateService;
 import com.cmvr.test.flow.runtime.operator.edge.EdgeOperateService;
+import com.cmvr.test.flow.runtime.operator.edge.VisionLocateOperateService;
 import com.cmvr.test.flow.runtime.operator.llm.InspectionMeterRecognizeOperateService;
 import com.cmvr.test.flow.runtime.operator.llm.LLMOperateService;
 import com.cmvr.test.model.vo.FlowActionRequestVO;
@@ -52,6 +53,7 @@ public class FlowActionExecutorService {
     private final EdgeArmOperateService edgeArmOperateService;
     private final EdgeManualInspectionOperateService edgeManualInspectionOperateService;
     private final EdgeDeviceCommandOperateService edgeDeviceCommandOperateService;
+    private final VisionLocateOperateService visionLocateOperateService;
     private final InspectionMeterRecognizeOperateService inspectionMeterRecognizeOperateService;
     private final List<EdgeOperateService> edgeOperateServices;
     private final List<LLMOperateService> llmOperateServices;
@@ -150,6 +152,14 @@ public class FlowActionExecutorService {
         String robotId = payload.getString("robotId");
         String deviceId = payload.getString("deviceId");
 
+        // Keep the legacy single-node fallback complete for actions introduced
+        // after the generic EDGE operator list was added.
+        if (ActionEnum.VISION_LOCATE_TARGET.equals(action)) {
+            TaskNodeExecuteMessage message = buildSingleNodeMessage(action, payload);
+            message.setRobotId(robotId);
+            return resultToString(visionLocateOperateService.execute(message));
+        }
+
         // 巡检报警监听组件只需要机器人 ID 和事件类型，不需要设备 ID。
         if (ActionEnum.INSPECTION_ALERT_LISTEN_START.equals(action)
                 || ActionEnum.INSPECTION_ALERT_LISTEN_STOP.equals(action)) {
diff --git a/cmvr-iot-test/src/main/java/com/cmvr/test/vision/domain/VisionCalibrationProfile.java b/cmvr-iot-test/src/main/java/com/cmvr/test/vision/domain/VisionCalibrationProfile.java
new file mode 100644
index 0000000..f4a2cb4
--- /dev/null
+++ b/cmvr-iot-test/src/main/java/com/cmvr/test/vision/domain/VisionCalibrationProfile.java
@@ -0,0 +1,37 @@
+package com.cmvr.test.vision.domain;
+
+import com.baomidou.mybatisplus.annotation.IdType;
+import com.baomidou.mybatisplus.annotation.TableId;
+import com.baomidou.mybatisplus.annotation.TableName;
+import com.cmvr.common.core.domain.BaseEntity;
+import lombok.Data;
+import lombok.EqualsAndHashCode;
+
+import java.math.BigDecimal;
+import java.time.LocalDateTime;
+
+@Data
+@EqualsAndHashCode(callSuper = true)
+@TableName("te_vision_calibration_profile")
+public class VisionCalibrationProfile extends BaseEntity {
+    @TableId(type = IdType.ASSIGN_UUID)
+    private String id;
+    private String profileType;
+    private String profileKey;
+    private String profileName;
+    private Integer versionNo;
+    private String robotId;
+    private String armDeviceId;
+    private String cameraDeviceId;
+    private String toolCode;
+    private String status;
+    private Boolean forcedEnabled;
+    private String rawMatrixJson;
+    private String correctionJson;
+    private String effectiveMatrixJson;
+    private BigDecimal meanErrorMm;
+    private BigDecimal maxErrorMm;
+    private Integer sampleCount;
+    private String metadataJson;
+    private LocalDateTime verifiedTime;
+}
diff --git a/cmvr-iot-test/src/main/java/com/cmvr/test/vision/domain/VisionCalibrationRecord.java b/cmvr-iot-test/src/main/java/com/cmvr/test/vision/domain/VisionCalibrationRecord.java
new file mode 100644
index 0000000..e89c585
--- /dev/null
+++ b/cmvr-iot-test/src/main/java/com/cmvr/test/vision/domain/VisionCalibrationRecord.java
@@ -0,0 +1,32 @@
+package com.cmvr.test.vision.domain;
+
+import com.baomidou.mybatisplus.annotation.IdType;
+import com.baomidou.mybatisplus.annotation.TableId;
+import com.baomidou.mybatisplus.annotation.TableName;
+import com.cmvr.common.core.domain.BaseEntity;
+import lombok.Data;
+import lombok.EqualsAndHashCode;
+
+import java.time.LocalDateTime;
+
+@Data
+@EqualsAndHashCode(callSuper = true)
+@TableName("te_vision_calibration_record")
+public class VisionCalibrationRecord extends BaseEntity {
+    @TableId(type = IdType.ASSIGN_UUID)
+    private String id;
+    private String profileId;
+    private String sessionType;
+    private String robotId;
+    private String armDeviceId;
+    private String cameraDeviceId;
+    private String toolCode;
+    private String status;
+    private Integer progressStep;
+    private Integer totalSteps;
+    private String samplesJson;
+    private String resultJson;
+    private String errorMessage;
+    private LocalDateTime startedTime;
+    private LocalDateTime finishedTime;
+}
diff --git a/cmvr-iot-test/src/main/java/com/cmvr/test/vision/mapper/VisionCalibrationProfileMapper.java b/cmvr-iot-test/src/main/java/com/cmvr/test/vision/mapper/VisionCalibrationProfileMapper.java
new file mode 100644
index 0000000..5bd36b3
--- /dev/null
+++ b/cmvr-iot-test/src/main/java/com/cmvr/test/vision/mapper/VisionCalibrationProfileMapper.java
@@ -0,0 +1,7 @@
+package com.cmvr.test.vision.mapper;
+
+import com.baomidou.mybatisplus.core.mapper.BaseMapper;
+import com.cmvr.test.vision.domain.VisionCalibrationProfile;
+
+public interface VisionCalibrationProfileMapper extends BaseMapper<VisionCalibrationProfile> {
+}
diff --git a/cmvr-iot-test/src/main/java/com/cmvr/test/vision/mapper/VisionCalibrationRecordMapper.java b/cmvr-iot-test/src/main/java/com/cmvr/test/vision/mapper/VisionCalibrationRecordMapper.java
new file mode 100644
index 0000000..c988692
--- /dev/null
+++ b/cmvr-iot-test/src/main/java/com/cmvr/test/vision/mapper/VisionCalibrationRecordMapper.java
@@ -0,0 +1,7 @@
+package com.cmvr.test.vision.mapper;
+
+import com.baomidou.mybatisplus.core.mapper.BaseMapper;
+import com.cmvr.test.vision.domain.VisionCalibrationRecord;
+
+public interface VisionCalibrationRecordMapper extends BaseMapper<VisionCalibrationRecord> {
+}
diff --git a/cmvr-iot-test/src/main/java/com/cmvr/test/vision/marker/MarkerObservation.java b/cmvr-iot-test/src/main/java/com/cmvr/test/vision/marker/MarkerObservation.java
new file mode 100644
index 0000000..0e08638
--- /dev/null
+++ b/cmvr-iot-test/src/main/java/com/cmvr/test/vision/marker/MarkerObservation.java
@@ -0,0 +1,8 @@
+package com.cmvr.test.vision.marker;
+
+import java.util.List;
+
+public record MarkerObservation(String imageUrl, double[] centerCamera, double[] normalCamera,
+                                double[] xAxisCamera, List<double[]> pixelCorners,
+                                int width, int height, double planeRmseMm) {
+}
diff --git a/cmvr-iot-test/src/main/java/com/cmvr/test/vision/marker/SquareMarkerDetector.java b/cmvr-iot-test/src/main/java/com/cmvr/test/vision/marker/SquareMarkerDetector.java
new file mode 100644
index 0000000..eaaa89a
--- /dev/null
+++ b/cmvr-iot-test/src/main/java/com/cmvr/test/vision/marker/SquareMarkerDetector.java
@@ -0,0 +1,287 @@
+package com.cmvr.test.vision.marker;
+
+import cn.hutool.http.HttpRequest;
+import cn.hutool.http.HttpResponse;
+import cmvr.api.CameraCommand;
+import com.cmvr.common.exception.GlobalException;
+import com.cmvr.edge.client.model.camera.EdgeCameraRgbdSnapshot;
+import nu.pattern.OpenCV;
+import org.opencv.core.Core;
+import org.opencv.core.CvType;
+import org.opencv.core.Mat;
+import org.opencv.core.MatOfByte;
+import org.opencv.core.MatOfDouble;
+import org.opencv.core.MatOfPoint;
+import org.opencv.core.MatOfPoint2f;
+import org.opencv.core.Point;
+import org.opencv.core.Rect;
+import org.opencv.core.Scalar;
+import org.opencv.imgcodecs.Imgcodecs;
+import org.opencv.imgproc.Imgproc;
+import org.springframework.stereotype.Component;
+
+import java.nio.ByteBuffer;
+import java.nio.ByteOrder;
+import java.util.ArrayList;
+import java.util.Arrays;
+import java.util.Comparator;
+import java.util.List;
+
+@Component
+public class SquareMarkerDetector {
+    static {
+        OpenCV.loadLocally();
+    }
+
+    public MarkerObservation detect(EdgeCameraRgbdSnapshot snapshot) {
+        Mat image = load(snapshot.getColorImageUrl());
+        try {
+            Point[] corners = findMarker(image);
+            Plane plane = fitDepthPlane(snapshot, corners);
+            Point centerPixel = average(corners);
+            double[] center = intersectPixel(snapshot, centerPixel, plane);
+            double[] topLeft = intersectPixel(snapshot, corners[0], plane);
+            double[] topRight = intersectPixel(snapshot, corners[1], plane);
+            double[] xAxis = normalize(subtract(topRight, topLeft));
+            return new MarkerObservation(snapshot.getColorImageUrl(), center, plane.normal, xAxis,
+                    Arrays.stream(corners).map(point -> new double[]{point.x, point.y}).toList(),
+                    snapshot.getWidth(), snapshot.getHeight(), plane.rmse * 1000.0);
+        } finally {
+            image.release();
+        }
+    }
+
+    private Mat load(String imageUrl) {
+        try (HttpResponse response = HttpRequest.get(imageUrl).timeout(10_000).execute()) {
+            if (!response.isOk()) throw new GlobalException("下载标定图片失败(" + response.getStatus() + ")");
+            MatOfByte bytes = new MatOfByte(response.bodyBytes());
+            try {
+                Mat image = Imgcodecs.imdecode(bytes, Imgcodecs.IMREAD_COLOR);
+                if (image.empty()) throw new GlobalException("标定图片无法解码");
+                return image;
+            } finally {
+                bytes.release();
+            }
+        } catch (GlobalException exception) {
+            throw exception;
+        } catch (RuntimeException exception) {
+            throw new GlobalException("读取标定图片失败: " + exception.getMessage());
+        }
+    }
+
+    private Point[] findMarker(Mat image) {
+        Mat gray = new Mat();
+        Mat binary = new Mat();
+        Mat hierarchy = new Mat();
+        List<MatOfPoint> contours = new ArrayList<>();
+        try {
+            Imgproc.cvtColor(image, gray, Imgproc.COLOR_BGR2GRAY);
+            Imgproc.GaussianBlur(gray, gray, new org.opencv.core.Size(3, 3), 0);
+            Imgproc.adaptiveThreshold(gray, binary, 255, Imgproc.ADAPTIVE_THRESH_GAUSSIAN_C,
+                    Imgproc.THRESH_BINARY, 21, 3);
+            Imgproc.findContours(binary, contours, hierarchy, Imgproc.RETR_TREE, Imgproc.CHAIN_APPROX_SIMPLE);
+            double imageArea = image.width() * (double) image.height();
+            Candidate best = null;
+            for (int index = 0; index < contours.size(); index++) {
+                MatOfPoint contour = contours.get(index);
+                double area = Math.abs(Imgproc.contourArea(contour));
+                if (area < 55 || area > imageArea * 0.65) continue;
+                MatOfPoint2f curve = new MatOfPoint2f(contour.toArray());
+                MatOfPoint2f approximation = new MatOfPoint2f();
+                try {
+                    double perimeter = Imgproc.arcLength(curve, true);
+                    Imgproc.approxPolyDP(curve, approximation, perimeter * 0.02, true);
+                    Point[] points = approximation.toArray();
+                    if (points.length != 4) continue;
+                    MatOfPoint polygon = new MatOfPoint(points);
+                    boolean convex = Imgproc.isContourConvex(polygon);
+                    polygon.release();
+                    if (!convex || !reasonableShape(points)) continue;
+                    Point[] ordered = order(points);
+                    double score = markerScore(gray, ordered) + nestedDepth(hierarchy, index) * 20 + Math.log(area);
+                    if (score > 25 && (best == null || score > best.score)) best = new Candidate(ordered, score);
+                } finally {
+                    curve.release();
+                    approximation.release();
+                }
+            }
+            if (best == null) throw new GlobalException("未识别到唯一方形标定码，请调整距离、角度或光线");
+            return best.corners;
+        } finally {
+            gray.release(); binary.release(); hierarchy.release(); contours.forEach(Mat::release);
+        }
+    }
+
+    private int nestedDepth(Mat hierarchy, int index) {
+        if (hierarchy.empty()) return 0;
+        int depth = 0;
+        int child = (int) hierarchy.get(0, index)[2];
+        while (child >= 0 && depth < 5) {
+            depth++;
+            child = (int) hierarchy.get(0, child)[2];
+        }
+        return depth;
+    }
+
+    private double markerScore(Mat gray, Point[] corners) {
+        Mat source = new MatOfPoint2f(corners);
+        Mat destination = new MatOfPoint2f(new Point(0, 0), new Point(159, 0), new Point(159, 159), new Point(0, 159));
+        Mat transform = Imgproc.getPerspectiveTransform(source, destination);
+        Mat warped = new Mat(160, 160, CvType.CV_8UC1);
+        try {
+            Imgproc.warpPerspective(gray, warped, transform, warped.size(), Imgproc.INTER_LINEAR);
+            Mat inner = warped.submat(new Rect(20, 20, 120, 120));
+            Mat center = warped.submat(new Rect(35, 35, 90, 90));
+            try {
+                Scalar mean = Core.mean(inner);
+                Scalar centerMean = Core.mean(center);
+                MatOfDouble meanMat = new MatOfDouble();
+                MatOfDouble std = new MatOfDouble();
+                Core.meanStdDev(center, meanMat, std);
+                double deviation = std.get(0, 0)[0];
+                meanMat.release(); std.release();
+                return deviation + Math.abs(mean.val[0] - centerMean.val[0]) * 0.2;
+            } finally {
+                inner.release(); center.release();
+            }
+        } finally {
+            source.release(); destination.release(); transform.release(); warped.release();
+        }
+    }
+
+    private Plane fitDepthPlane(EdgeCameraRgbdSnapshot snapshot, Point[] corners) {
+        List<double[]> points = new ArrayList<>();
+        for (int row = 1; row <= 11; row++) {
+            double v = row / 12.0;
+            for (int column = 1; column <= 11; column++) {
+                double u = column / 12.0;
+                Point pixel = bilinear(corners, u, v);
+                double depth = medianDepth(snapshot, (int) Math.round(pixel.x), (int) Math.round(pixel.y), 2);
+                if (Double.isFinite(depth)) points.add(backProject(snapshot, pixel.x, pixel.y, depth));
+            }
+        }
+        if (points.size() < 30) throw new GlobalException("标定码区域有效深度点不足");
+        if (points.size() < 30) {
+            Point center = average(corners);
+            double centerDepth = medianDepth(snapshot, (int) Math.round(center.x), (int) Math.round(center.y), 8);
+            if (Double.isFinite(centerDepth)) {
+                points.clear();
+                points.add(backProject(snapshot, center.x, center.y, centerDepth));
+                double[] normal = {0, 0, -1};
+                return new Plane(normal, centerDepth, 0.0);
+            }
+        }
+        double medianZ = points.stream().mapToDouble(point -> point[2]).sorted().skip(points.size() / 2).findFirst().orElseThrow();
+        points.removeIf(point -> Math.abs(point[2] - medianZ) > 0.025);
+        double[][] design = new double[3][3];
+        double[] target = new double[3];
+        for (double[] point : points) {
+            double[] row = {point[0], point[1], 1};
+            for (int i = 0; i < 3; i++) {
+                target[i] += row[i] * point[2];
+                for (int j = 0; j < 3; j++) design[i][j] += row[i] * row[j];
+            }
+        }
+        double[] solution = solve3(design, target);
+        double[] normal = normalize(new double[]{solution[0], solution[1], -1});
+        if (dot(normal, new double[]{0, 0, 1}) > 0) normal = scale(normal, -1);
+        double d = solution[2] / Math.sqrt(solution[0] * solution[0] + solution[1] * solution[1] + 1);
+        double squared = 0;
+        for (double[] point : points) {
+            double error = dot(normal, point) + d;
+            squared += error * error;
+        }
+        return new Plane(normal, d, Math.sqrt(squared / points.size()));
+    }
+
+    private double[] intersectPixel(EdgeCameraRgbdSnapshot snapshot, Point pixel, Plane plane) {
+        double[] ray = {(pixel.x - snapshot.getCx()) / snapshot.getFx(),
+                (pixel.y - snapshot.getCy()) / snapshot.getFy(), 1};
+        double scale = -plane.d / dot(plane.normal, ray);
+        return scale(ray, scale);
+    }
+
+    private double medianDepth(EdgeCameraRgbdSnapshot snapshot, int x, int y, int radius) {
+        List<Double> values = new ArrayList<>();
+        for (int py = Math.max(0, y - radius); py <= Math.min(snapshot.getHeight() - 1, y + radius); py++)
+            for (int px = Math.max(0, x - radius); px <= Math.min(snapshot.getWidth() - 1, x + radius); px++) {
+                double raw = readDepth(snapshot, py * snapshot.getWidth() + px);
+                if (Double.isFinite(raw) && raw > 0) values.add(raw * snapshot.getDepthUnitScale());
+            }
+        if (values.isEmpty()) return Double.NaN;
+        values.sort(Double::compareTo);
+        return values.get(values.size() / 2);
+    }
+
+    private double readDepth(EdgeCameraRgbdSnapshot snapshot, int index) {
+        byte[] bytes = snapshot.getDepthData();
+        CameraCommand.FrameData.FrameType type = snapshot.getDepthType();
+        if (type == CameraCommand.FrameData.FrameType.U8C1) return Byte.toUnsignedInt(bytes[index]);
+        ByteBuffer buffer = ByteBuffer.wrap(bytes).order(ByteOrder.LITTLE_ENDIAN);
+        if (type == CameraCommand.FrameData.FrameType.F32C1) return buffer.getFloat(index * Float.BYTES);
+        if (type == CameraCommand.FrameData.FrameType.U16C1 || type == CameraCommand.FrameData.FrameType.F16C1)
+            return Short.toUnsignedInt(buffer.getShort(index * Short.BYTES));
+        return Double.NaN;
+    }
+
+    private double[] backProject(EdgeCameraRgbdSnapshot snapshot, double x, double y, double depth) {
+        return new double[]{(x - snapshot.getCx()) * depth / snapshot.getFx(),
+                (y - snapshot.getCy()) * depth / snapshot.getFy(), depth};
+    }
+
+    private boolean reasonableShape(Point[] points) {
+        Point[] ordered = order(points);
+        double[] lengths = new double[4];
+        for (int index = 0; index < 4; index++) lengths[index] = distance(ordered[index], ordered[(index + 1) % 4]);
+        double min = Arrays.stream(lengths).min().orElse(0), max = Arrays.stream(lengths).max().orElse(1);
+        return min > 8 && max / min < 2.3;
+    }
+
+    private Point[] order(Point[] points) {
+        Point[] ordered = new Point[4];
+        ordered[0] = Arrays.stream(points).min(Comparator.comparingDouble(point -> point.x + point.y)).orElseThrow();
+        ordered[2] = Arrays.stream(points).max(Comparator.comparingDouble(point -> point.x + point.y)).orElseThrow();
+        ordered[1] = Arrays.stream(points).max(Comparator.comparingDouble(point -> point.x - point.y)).orElseThrow();
+        ordered[3] = Arrays.stream(points).min(Comparator.comparingDouble(point -> point.x - point.y)).orElseThrow();
+        return ordered;
+    }
+
+    private Point bilinear(Point[] corners, double u, double v) {
+        return new Point((1-u)*(1-v)*corners[0].x + u*(1-v)*corners[1].x + u*v*corners[2].x + (1-u)*v*corners[3].x,
+                (1-u)*(1-v)*corners[0].y + u*(1-v)*corners[1].y + u*v*corners[2].y + (1-u)*v*corners[3].y);
+    }
+
+    private Point average(Point[] points) {
+        return new Point(Arrays.stream(points).mapToDouble(point -> point.x).average().orElse(0),
+                Arrays.stream(points).mapToDouble(point -> point.y).average().orElse(0));
+    }
+
+    private double[] solve3(double[][] matrix, double[] right) {
+        double[][] a = new double[3][4];
+        for (int row = 0; row < 3; row++) { System.arraycopy(matrix[row], 0, a[row], 0, 3); a[row][3] = right[row]; }
+        for (int column = 0; column < 3; column++) {
+            int pivot = column;
+            for (int row = column + 1; row < 3; row++) if (Math.abs(a[row][column]) > Math.abs(a[pivot][column])) pivot = row;
+            double[] temp = a[column]; a[column] = a[pivot]; a[pivot] = temp;
+            if (Math.abs(a[column][column]) < 1e-12) throw new GlobalException("标定码平面拟合失败");
+            for (int row = column + 1; row < 3; row++) {
+                double factor = a[row][column] / a[column][column];
+                for (int index = column; index < 4; index++) a[row][index] -= factor * a[column][index];
+            }
+        }
+        double[] result = new double[3];
+        for (int row = 2; row >= 0; row--) {
+            double value = a[row][3]; for (int column = row + 1; column < 3; column++) value -= a[row][column] * result[column];
+            result[row] = value / a[row][row];
+        }
+        return result;
+    }
+
+    private double distance(Point left, Point right) { return Math.hypot(left.x - right.x, left.y - right.y); }
+    private double[] subtract(double[] left, double[] right) { return new double[]{left[0]-right[0], left[1]-right[1], left[2]-right[2]}; }
+    private double dot(double[] left, double[] right) { return left[0]*right[0] + left[1]*right[1] + left[2]*right[2]; }
+    private double[] scale(double[] value, double factor) { return new double[]{value[0]*factor, value[1]*factor, value[2]*factor}; }
+    private double[] normalize(double[] value) { double length = Math.sqrt(dot(value, value)); if (length < 1e-10) throw new GlobalException("标定方向无效"); return scale(value, 1/length); }
+    private record Candidate(Point[] corners, double score) { }
+    private record Plane(double[] normal, double d, double rmse) { }
+}
diff --git a/cmvr-iot-test/src/main/java/com/cmvr/test/vision/math/CalibrationResult.java b/cmvr-iot-test/src/main/java/com/cmvr/test/vision/math/CalibrationResult.java
new file mode 100644
index 0000000..72dd67e
--- /dev/null
+++ b/cmvr-iot-test/src/main/java/com/cmvr/test/vision/math/CalibrationResult.java
@@ -0,0 +1,5 @@
+package com.cmvr.test.vision.math;
+
+public record CalibrationResult(double[][] matrix, double meanErrorMm, double maxErrorMm,
+                                int sampleCount, double[] fixedPoint) {
+}
diff --git a/cmvr-iot-test/src/main/java/com/cmvr/test/vision/math/CalibrationSample.java b/cmvr-iot-test/src/main/java/com/cmvr/test/vision/math/CalibrationSample.java
new file mode 100644
index 0000000..5aa8e70
--- /dev/null
+++ b/cmvr-iot-test/src/main/java/com/cmvr/test/vision/math/CalibrationSample.java
@@ -0,0 +1,4 @@
+package com.cmvr.test.vision.math;
+
+public record CalibrationSample(double[][] baseToFlange, double[] cameraPoint) {
+}
diff --git a/cmvr-iot-test/src/main/java/com/cmvr/test/vision/math/VisionCalibrationMath.java b/cmvr-iot-test/src/main/java/com/cmvr/test/vision/math/VisionCalibrationMath.java
new file mode 100644
index 0000000..213169f
--- /dev/null
+++ b/cmvr-iot-test/src/main/java/com/cmvr/test/vision/math/VisionCalibrationMath.java
@@ -0,0 +1,355 @@
+package com.cmvr.test.vision.math;
+
+import cmvr.api.ArmCommand;
+import com.alibaba.fastjson2.JSON;
+import com.alibaba.fastjson2.JSONArray;
+import com.cmvr.common.exception.GlobalException;
+import com.cmvr.test.vision.model.VisionCalibrationRequests;
+
+import java.util.ArrayList;
+import java.util.Arrays;
+import java.util.List;
+
+public final class VisionCalibrationMath {
+    private VisionCalibrationMath() {
+    }
+
+    public static CalibrationResult solveHandEye(List<CalibrationSample> samples, double[][] initial) {
+        if (samples == null || samples.size() < 8) {
+            throw new GlobalException("手眼标定至少需要8组有效样本");
+        }
+        double[] parameters = new double[9];
+        double[][] seed = initial == null ? identity() : initial;
+        double[] rotation = rotationVector(seed);
+        System.arraycopy(rotation, 0, parameters, 0, 3);
+        parameters[3] = seed[0][3];
+        parameters[4] = seed[1][3];
+        parameters[5] = seed[2][3];
+        double[] firstPrediction = transform(multiply(samples.get(0).baseToFlange(), seed),
+                samples.get(0).cameraPoint());
+        System.arraycopy(firstPrediction, 0, parameters, 6, 3);
+
+        double lambda = 1e-5;
+        double currentCost = cost(samples, parameters);
+        for (int iteration = 0; iteration < 120; iteration++) {
+            double[] residual = residuals(samples, parameters);
+            double[][] jacobian = numericJacobian(samples, parameters, residual);
+            double[][] normal = new double[parameters.length][parameters.length];
+            double[] rhs = new double[parameters.length];
+            for (int row = 0; row < residual.length; row++) {
+                for (int column = 0; column < parameters.length; column++) {
+                    rhs[column] -= jacobian[row][column] * residual[row];
+                    for (int other = 0; other < parameters.length; other++) {
+                        normal[column][other] += jacobian[row][column] * jacobian[row][other];
+                    }
+                }
+            }
+            for (int index = 0; index < parameters.length; index++) normal[index][index] += lambda;
+            double[] delta = solve(normal, rhs);
+            double[] candidate = parameters.clone();
+            for (int index = 0; index < candidate.length; index++) candidate[index] += delta[index];
+            double candidateCost = cost(samples, candidate);
+            if (candidateCost < currentCost) {
+                parameters = candidate;
+                if (norm(delta) < 1e-10 || Math.abs(currentCost - candidateCost) < 1e-14) break;
+                currentCost = candidateCost;
+                lambda = Math.max(1e-10, lambda * 0.35);
+            } else {
+                lambda = Math.min(1e6, lambda * 8.0);
+            }
+        }
+        double[][] matrix = fromRotationVector(Arrays.copyOf(parameters, 3));
+        matrix[0][3] = parameters[3];
+        matrix[1][3] = parameters[4];
+        matrix[2][3] = parameters[5];
+        return result(matrix, samples, Arrays.copyOfRange(parameters, 6, 9));
+    }
+
+    public static CalibrationResult solveToolPivot(List<double[][]> baseToFlangeSamples) {
+        if (baseToFlangeSamples == null || baseToFlangeSamples.size() < 5) {
+            throw new GlobalException("工具TCP标定至少需要5个不同姿态");
+        }
+        int rows = baseToFlangeSamples.size() * 3;
+        double[][] design = new double[rows][6];
+        double[] target = new double[rows];
+        for (int sampleIndex = 0; sampleIndex < baseToFlangeSamples.size(); sampleIndex++) {
+            double[][] pose = baseToFlangeSamples.get(sampleIndex);
+            for (int axis = 0; axis < 3; axis++) {
+                int row = sampleIndex * 3 + axis;
+                System.arraycopy(pose[axis], 0, design[row], 0, 3);
+                design[row][3 + axis] = -1.0;
+                target[row] = -pose[axis][3];
+            }
+        }
+        double[] solution = leastSquares(design, target);
+        double[] tcp = Arrays.copyOf(solution, 3);
+        double[] fixed = Arrays.copyOfRange(solution, 3, 6);
+        double[][] matrix = identity();
+        matrix[0][3] = tcp[0];
+        matrix[1][3] = tcp[1];
+        matrix[2][3] = tcp[2];
+        List<Double> errors = new ArrayList<>();
+        for (double[][] pose : baseToFlangeSamples) {
+            errors.add(distance(transform(pose, tcp), fixed) * 1000.0);
+        }
+        return new CalibrationResult(matrix, mean(errors), errors.stream().mapToDouble(Double::doubleValue).max().orElse(0),
+                baseToFlangeSamples.size(), fixed);
+    }
+
+    public static double[][] applyToolDirection(double[][] pivotMatrix, double[][] baseToFlange,
+                                                 double[] planeNormalBase, double[] planeXBase) {
+        double[][] flangeToBaseRotation = transposeRotation(baseToFlange);
+        double[] z = normalize(rotationTransform(flangeToBaseRotation, scale(normalize(planeNormalBase), -1)));
+        double[] xSeed = normalize(rotationTransform(flangeToBaseRotation, normalize(planeXBase)));
+        double[] x = normalize(subtract(xSeed, scale(z, dot(xSeed, z))));
+        if (norm(x) < 1e-8) x = perpendicular(z);
+        double[] y = normalize(cross(z, x));
+        x = normalize(cross(y, z));
+        double[][] result = copy(pivotMatrix);
+        setColumn(result, 0, x);
+        setColumn(result, 1, y);
+        setColumn(result, 2, z);
+        return result;
+    }
+
+    public static double[][] applyCorrection(double[][] raw, VisionCalibrationRequests.Correction correction) {
+        if (correction == null) return copy(raw);
+        double[][] delta = fromEuler(Math.toRadians(value(correction.getRxDeg())),
+                Math.toRadians(value(correction.getRyDeg())), Math.toRadians(value(correction.getRzDeg())));
+        delta[0][3] = value(correction.getXMm()) / 1000.0;
+        delta[1][3] = value(correction.getYMm()) / 1000.0;
+        delta[2][3] = value(correction.getZMm()) / 1000.0;
+        return multiply(raw, delta);
+    }
+
+    public static double[][] fromProto(ArmCommand.TransformMatrix4x4 matrix) {
+        return new double[][]{
+                {matrix.getM00(), matrix.getM01(), matrix.getM02(), matrix.getM03()},
+                {matrix.getM10(), matrix.getM11(), matrix.getM12(), matrix.getM13()},
+                {matrix.getM20(), matrix.getM21(), matrix.getM22(), matrix.getM23()},
+                {matrix.getM30(), matrix.getM31(), matrix.getM32(), matrix.getM33()}
+        };
+    }
+
+    /** Converts the arm getPose response (metres/radians) to a homogeneous pose matrix. */
+    public static double[][] fromCartesianPose(ArmCommand.CartesianPose pose) {
+        if (pose == null) throw new GlobalException("机械臂位姿为空");
+        double[][] matrix = fromEuler(pose.getRx(), pose.getRy(), pose.getRz());
+        matrix[0][3] = pose.getX();
+        matrix[1][3] = pose.getY();
+        matrix[2][3] = pose.getZ();
+        return matrix;
+    }
+
+    public static String toJson(double[][] matrix) {
+        JSONArray values = new JSONArray();
+        for (double[] row : matrix) for (double value : row) values.add(value);
+        return values.toJSONString();
+    }
+
+    public static double[][] parseMatrix(String json) {
+        try {
+            JSONArray values = JSON.parseArray(json);
+            if (values == null || values.size() != 16) throw new IllegalArgumentException();
+            double[][] matrix = new double[4][4];
+            for (int index = 0; index < 16; index++) matrix[index / 4][index % 4] = values.getDoubleValue(index);
+            return matrix;
+        } catch (RuntimeException exception) {
+            throw new GlobalException("标定矩阵格式无效");
+        }
+    }
+
+    public static double[][] multiply(double[][] left, double[][] right) {
+        double[][] result = new double[4][4];
+        for (int row = 0; row < 4; row++) for (int column = 0; column < 4; column++)
+            for (int index = 0; index < 4; index++) result[row][column] += left[row][index] * right[index][column];
+        return result;
+    }
+
+    public static double[][] rigidInverse(double[][] matrix) {
+        double[][] result = identity();
+        for (int row = 0; row < 3; row++) for (int column = 0; column < 3; column++) result[row][column] = matrix[column][row];
+        for (int row = 0; row < 3; row++) result[row][3] = -(result[row][0] * matrix[0][3]
+                + result[row][1] * matrix[1][3] + result[row][2] * matrix[2][3]);
+        return result;
+    }
+
+    public static double[] transform(double[][] matrix, double[] point) {
+        return new double[]{matrix[0][0] * point[0] + matrix[0][1] * point[1] + matrix[0][2] * point[2] + matrix[0][3],
+                matrix[1][0] * point[0] + matrix[1][1] * point[1] + matrix[1][2] * point[2] + matrix[1][3],
+                matrix[2][0] * point[0] + matrix[2][1] * point[1] + matrix[2][2] * point[2] + matrix[2][3]};
+    }
+
+    public static double[] pose(double[][] matrix) {
+        double ry = Math.asin(clamp(-matrix[2][0], -1, 1));
+        double rx = Math.abs(Math.cos(ry)) > 1e-8 ? Math.atan2(matrix[2][1], matrix[2][2])
+                : Math.atan2(-matrix[1][2], matrix[1][1]);
+        double rz = Math.abs(Math.cos(ry)) > 1e-8 ? Math.atan2(matrix[1][0], matrix[0][0]) : 0;
+        return new double[]{matrix[0][3], matrix[1][3], matrix[2][3], rx, ry, rz};
+    }
+
+    public static double[][] fromEuler(double rx, double ry, double rz) {
+        double cx = Math.cos(rx), sx = Math.sin(rx), cy = Math.cos(ry), sy = Math.sin(ry), cz = Math.cos(rz), sz = Math.sin(rz);
+        double[][] result = identity();
+        result[0][0] = cz * cy; result[0][1] = cz * sy * sx - sz * cx; result[0][2] = cz * sy * cx + sz * sx;
+        result[1][0] = sz * cy; result[1][1] = sz * sy * sx + cz * cx; result[1][2] = sz * sy * cx - cz * sx;
+        result[2][0] = -sy; result[2][1] = cy * sx; result[2][2] = cy * cx;
+        return result;
+    }
+
+    public static double[][] identity() {
+        double[][] result = new double[4][4];
+        for (int index = 0; index < 4; index++) result[index][index] = 1;
+        return result;
+    }
+
+    private static CalibrationResult result(double[][] matrix, List<CalibrationSample> samples, double[] fixed) {
+        List<Double> errors = new ArrayList<>();
+        for (CalibrationSample sample : samples) {
+            double[] prediction = transform(multiply(sample.baseToFlange(), matrix), sample.cameraPoint());
+            errors.add(distance(prediction, fixed) * 1000.0);
+        }
+        return new CalibrationResult(matrix, mean(errors), errors.stream().mapToDouble(Double::doubleValue).max().orElse(0),
+                samples.size(), fixed);
+    }
+
+    private static double[] residuals(List<CalibrationSample> samples, double[] parameters) {
+        double[][] rotation = fromRotationVector(Arrays.copyOf(parameters, 3));
+        rotation[0][3] = parameters[3]; rotation[1][3] = parameters[4]; rotation[2][3] = parameters[5];
+        double[] fixed = Arrays.copyOfRange(parameters, 6, 9);
+        double[] result = new double[samples.size() * 3];
+        for (int index = 0; index < samples.size(); index++) {
+            CalibrationSample sample = samples.get(index);
+            double[] predicted = transform(multiply(sample.baseToFlange(), rotation), sample.cameraPoint());
+            for (int axis = 0; axis < 3; axis++) result[index * 3 + axis] = predicted[axis] - fixed[axis];
+        }
+        return result;
+    }
+
+    private static double[][] numericJacobian(List<CalibrationSample> samples, double[] parameters, double[] residual) {
+        double[][] result = new double[residual.length][parameters.length];
+        for (int column = 0; column < parameters.length; column++) {
+            double[] shifted = parameters.clone();
+            double epsilon = column < 3 ? 1e-6 : 1e-5;
+            shifted[column] += epsilon;
+            double[] shiftedResidual = residuals(samples, shifted);
+            for (int row = 0; row < residual.length; row++) result[row][column] = (shiftedResidual[row] - residual[row]) / epsilon;
+        }
+        return result;
+    }
+
+    private static double cost(List<CalibrationSample> samples, double[] parameters) {
+        double sum = 0;
+        for (double value : residuals(samples, parameters)) sum += value * value;
+        return sum;
+    }
+
+    private static double[][] fromRotationVector(double[] vector) {
+        double angle = norm(vector);
+        if (angle < 1e-12) return identity();
+        double x = vector[0] / angle, y = vector[1] / angle, z = vector[2] / angle;
+        double c = Math.cos(angle), s = Math.sin(angle), t = 1 - c;
+        double[][] result = identity();
+        result[0][0] = t*x*x+c; result[0][1] = t*x*y-s*z; result[0][2] = t*x*z+s*y;
+        result[1][0] = t*x*y+s*z; result[1][1] = t*y*y+c; result[1][2] = t*y*z-s*x;
+        result[2][0] = t*x*z-s*y; result[2][1] = t*y*z+s*x; result[2][2] = t*z*z+c;
+        return result;
+    }
+
+    private static double[] rotationVector(double[][] matrix) {
+        double angle = Math.acos(clamp((matrix[0][0] + matrix[1][1] + matrix[2][2] - 1) / 2, -1, 1));
+        if (angle < 1e-10) return new double[3];
+        double scale = angle / (2 * Math.sin(angle));
+        return new double[]{(matrix[2][1] - matrix[1][2]) * scale,
+                (matrix[0][2] - matrix[2][0]) * scale, (matrix[1][0] - matrix[0][1]) * scale};
+    }
+
+    private static double[] leastSquares(double[][] design, double[] target) {
+        int columns = design[0].length;
+        double[][] normal = new double[columns][columns];
+        double[] rhs = new double[columns];
+        for (int row = 0; row < design.length; row++) for (int column = 0; column < columns; column++) {
+            rhs[column] += design[row][column] * target[row];
+            for (int other = 0; other < columns; other++) normal[column][other] += design[row][column] * design[row][other];
+        }
+        for (int index = 0; index < columns; index++) normal[index][index] += 1e-12;
+        return solve(normal, rhs);
+    }
+
+    private static double[] solve(double[][] source, double[] sourceRight) {
+        int size = sourceRight.length;
+        double[][] matrix = new double[size][size];
+        for (int row = 0; row < size; row++) matrix[row] = source[row].clone();
+        double[] right = sourceRight.clone();
+        for (int column = 0; column < size; column++) {
+            int pivot = column;
+            for (int row = column + 1; row < size; row++) if (Math.abs(matrix[row][column]) > Math.abs(matrix[pivot][column])) pivot = row;
+            if (Math.abs(matrix[pivot][column]) < 1e-14) throw new GlobalException("标定姿态变化不足，无法求解");
+            double[] temp = matrix[column]; matrix[column] = matrix[pivot]; matrix[pivot] = temp;
+            double tempValue = right[column]; right[column] = right[pivot]; right[pivot] = tempValue;
+            for (int row = column + 1; row < size; row++) {
+                double factor = matrix[row][column] / matrix[column][column];
+                for (int index = column; index < size; index++) matrix[row][index] -= factor * matrix[column][index];
+                right[row] -= factor * right[column];
+            }
+        }
+        double[] result = new double[size];
+        for (int row = size - 1; row >= 0; row--) {
+            double value = right[row];
+            for (int column = row + 1; column < size; column++) value -= matrix[row][column] * result[column];
+            result[row] = value / matrix[row][row];
+        }
+        return result;
+    }
+
+    private static double[][] transposeRotation(double[][] matrix) {
+        double[][] result = identity();
+        for (int row = 0; row < 3; row++) for (int column = 0; column < 3; column++) result[row][column] = matrix[column][row];
+        return result;
+    }
+
+    private static double[] rotationTransform(double[][] rotation, double[] value) {
+        return new double[]{dot(Arrays.copyOf(rotation[0], 3), value), dot(Arrays.copyOf(rotation[1], 3), value), dot(Arrays.copyOf(rotation[2], 3), value)};
+    }
+
+    private static void setColumn(double[][] matrix, int column, double[] value) {
+        for (int row = 0; row < 3; row++) matrix[row][column] = value[row];
+    }
+
+    private static double[][] copy(double[][] matrix) {
+        double[][] result = new double[matrix.length][];
+        for (int index = 0; index < matrix.length; index++) result[index] = matrix[index].clone();
+        return result;
+    }
+
+    private static double[] perpendicular(double[] vector) {
+        double[] seed = Math.abs(vector[0]) < 0.8 ? new double[]{1, 0, 0} : new double[]{0, 1, 0};
+        return normalize(cross(seed, vector));
+    }
+
+    private static double[] cross(double[] left, double[] right) {
+        return new double[]{left[1]*right[2]-left[2]*right[1], left[2]*right[0]-left[0]*right[2], left[0]*right[1]-left[1]*right[0]};
+    }
+
+    private static double dot(double[] left, double[] right) {
+        double result = 0; for (int index = 0; index < left.length; index++) result += left[index] * right[index]; return result;
+    }
+
+    private static double[] subtract(double[] left, double[] right) {
+        return new double[]{left[0]-right[0], left[1]-right[1], left[2]-right[2]};
+    }
+
+    private static double[] scale(double[] value, double factor) {
+        return new double[]{value[0]*factor, value[1]*factor, value[2]*factor};
+    }
+
+    private static double[] normalize(double[] value) {
+        double length = norm(value); if (length < 1e-12) throw new GlobalException("无法确定标定方向"); return scale(value, 1 / length);
+    }
+
+    private static double norm(double[] value) { double result = 0; for (double item : value) result += item * item; return Math.sqrt(result); }
+    private static double distance(double[] left, double[] right) { return norm(subtract(left, right)); }
+    private static double mean(List<Double> values) { return values.stream().mapToDouble(Double::doubleValue).average().orElse(0); }
+    private static double value(Double value) { return value == null ? 0 : value; }
+    private static double clamp(double value, double min, double max) { return Math.max(min, Math.min(max, value)); }
+}
diff --git a/cmvr-iot-test/src/main/java/com/cmvr/test/vision/model/VisionCalibrationRequests.java b/cmvr-iot-test/src/main/java/com/cmvr/test/vision/model/VisionCalibrationRequests.java
new file mode 100644
index 0000000..6bd5d16
--- /dev/null
+++ b/cmvr-iot-test/src/main/java/com/cmvr/test/vision/model/VisionCalibrationRequests.java
@@ -0,0 +1,46 @@
+package com.cmvr.test.vision.model;
+
+import lombok.Data;
+
+public final class VisionCalibrationRequests {
+    private VisionCalibrationRequests() {
+    }
+
+    @Data
+    public static class HandEyeStart {
+        private String profileName;
+        private String robotId;
+        private String armDeviceId;
+        private String cameraDeviceId;
+        private Double translationRangeMm = 10.0;
+        private Double rotationRangeDeg = 5.0;
+        private Double markerSizeMm = 20.0;
+        private Double velocity = 0.01;
+        private Double acceleration = 0.03;
+        private Integer settleTimeMs = 500;
+    }
+
+    @Data
+    public static class ToolStart {
+        private String profileName;
+        private String toolCode;
+        private String robotId;
+        private String armDeviceId;
+        private String cameraDeviceId;
+    }
+
+    @Data
+    public static class Correction {
+        private Double xMm = 0.0;
+        private Double yMm = 0.0;
+        private Double zMm = 0.0;
+        private Double rxDeg = 0.0;
+        private Double ryDeg = 0.0;
+        private Double rzDeg = 0.0;
+    }
+
+    @Data
+    public static class Activate {
+        private Boolean forced = false;
+    }
+}
diff --git a/cmvr-iot-test/src/main/java/com/cmvr/test/vision/model/VisionCalibrationSessionView.java b/cmvr-iot-test/src/main/java/com/cmvr/test/vision/model/VisionCalibrationSessionView.java
new file mode 100644
index 0000000..3d2cddb
--- /dev/null
+++ b/cmvr-iot-test/src/main/java/com/cmvr/test/vision/model/VisionCalibrationSessionView.java
@@ -0,0 +1,26 @@
+package com.cmvr.test.vision.model;
+
+import lombok.Builder;
+import lombok.Data;
+
+import java.util.List;
+
+@Data
+@Builder
+public class VisionCalibrationSessionView {
+    private String sessionId;
+    private String sessionType;
+    private String status;
+    private Integer currentStep;
+    private Integer totalSteps;
+    private Integer sampleCount;
+    private String message;
+    private String imageUrl;
+    private String profileId;
+    private Double meanErrorMm;
+    private Double maxErrorMm;
+    private Double directionErrorDeg;
+    private List<double[]> markerCorners;
+    private Integer imageWidth;
+    private Integer imageHeight;
+}
diff --git a/cmvr-iot-test/src/main/java/com/cmvr/test/vision/service/VisionCalibrationProfileService.java b/cmvr-iot-test/src/main/java/com/cmvr/test/vision/service/VisionCalibrationProfileService.java
new file mode 100644
index 0000000..950686d
--- /dev/null
+++ b/cmvr-iot-test/src/main/java/com/cmvr/test/vision/service/VisionCalibrationProfileService.java
@@ -0,0 +1,190 @@
+package com.cmvr.test.vision.service;
+
+import cn.hutool.core.util.StrUtil;
+import cn.hutool.crypto.digest.DigestUtil;
+import com.alibaba.fastjson2.JSON;
+import com.alibaba.fastjson2.JSONObject;
+import com.baomidou.mybatisplus.core.conditions.query.LambdaQueryWrapper;
+import com.baomidou.mybatisplus.core.conditions.update.LambdaUpdateWrapper;
+import com.baomidou.mybatisplus.spring.service.impl.ServiceImpl;
+import com.cmvr.common.exception.GlobalException;
+import com.cmvr.test.vision.domain.VisionCalibrationProfile;
+import com.cmvr.test.vision.mapper.VisionCalibrationProfileMapper;
+import com.cmvr.test.vision.math.CalibrationResult;
+import com.cmvr.test.vision.math.VisionCalibrationMath;
+import com.cmvr.test.vision.model.VisionCalibrationRequests;
+import org.springframework.stereotype.Service;
+import org.springframework.transaction.annotation.Transactional;
+
+import java.math.BigDecimal;
+import java.math.RoundingMode;
+import java.time.LocalDateTime;
+import java.util.List;
+
+@Service
+public class VisionCalibrationProfileService
+        extends ServiceImpl<VisionCalibrationProfileMapper, VisionCalibrationProfile> {
+
+    public static final String HAND_EYE = "HAND_EYE";
+    public static final String TOOL_TCP = "TOOL_TCP";
+    public static final String ACTIVE = "ACTIVE";
+    public static final String DRAFT = "DRAFT";
+    public static final String INVALID = "INVALID";
+
+    public List<VisionCalibrationProfile> list(String profileType, String robotId, String keyword) {
+        LambdaQueryWrapper<VisionCalibrationProfile> query = new LambdaQueryWrapper<VisionCalibrationProfile>()
+                .eq(StrUtil.isNotBlank(profileType), VisionCalibrationProfile::getProfileType, profileType)
+                .eq(StrUtil.isNotBlank(robotId), VisionCalibrationProfile::getRobotId, robotId)
+                .and(StrUtil.isNotBlank(keyword), wrapper -> wrapper
+                        .like(VisionCalibrationProfile::getProfileName, keyword)
+                        .or().like(VisionCalibrationProfile::getToolCode, keyword))
+                .orderByDesc(VisionCalibrationProfile::getCreateTime);
+        return baseMapper.selectList(query);
+    }
+
+    public VisionCalibrationProfile require(String id) {
+        VisionCalibrationProfile profile = baseMapper.selectById(id);
+        if (profile == null) throw new GlobalException("标定方案不存在");
+        return profile;
+    }
+
+    public VisionCalibrationProfile activeHandEye(String robotId, String armDeviceId, String cameraDeviceId) {
+        return active(profileKey(HAND_EYE, robotId, armDeviceId, cameraDeviceId));
+    }
+
+    public VisionCalibrationProfile activeTool(String robotId, String armDeviceId, String toolCode) {
+        if (StrUtil.isBlank(toolCode)) throw new GlobalException("请选择末端工具");
+        return active(profileKey(TOOL_TCP, robotId, armDeviceId, toolCode));
+    }
+
+    private VisionCalibrationProfile active(String key) {
+        VisionCalibrationProfile profile = baseMapper.selectOne(new LambdaQueryWrapper<VisionCalibrationProfile>()
+                .eq(VisionCalibrationProfile::getProfileKey, key)
+                .eq(VisionCalibrationProfile::getStatus, ACTIVE)
+                .orderByDesc(VisionCalibrationProfile::getVersionNo)
+                .last("limit 1"));
+        if (profile == null) throw new GlobalException("当前设备没有已启用的标定方案");
+        return profile;
+    }
+
+    @Transactional(rollbackFor = Exception.class)
+    public VisionCalibrationProfile createHandEye(VisionCalibrationRequests.HandEyeStart request,
+                                                   CalibrationResult result, JSONObject metadata,
+                                                   String operator) {
+        String key = profileKey(HAND_EYE, request.getRobotId(), request.getArmDeviceId(), request.getCameraDeviceId());
+        return create(key, HAND_EYE, request.getProfileName(), request.getRobotId(), request.getArmDeviceId(),
+                request.getCameraDeviceId(), null, result, metadata, operator);
+    }
+
+    @Transactional(rollbackFor = Exception.class)
+    public VisionCalibrationProfile createTool(VisionCalibrationRequests.ToolStart request,
+                                                CalibrationResult result, double[][] finalMatrix,
+                                                JSONObject metadata, String operator) {
+        String key = profileKey(TOOL_TCP, request.getRobotId(), request.getArmDeviceId(), request.getToolCode());
+        CalibrationResult complete = new CalibrationResult(finalMatrix, result.meanErrorMm(), result.maxErrorMm(),
+                result.sampleCount(), result.fixedPoint());
+        return create(key, TOOL_TCP, request.getProfileName(), request.getRobotId(), request.getArmDeviceId(),
+                request.getCameraDeviceId(), request.getToolCode(), complete, metadata, operator);
+    }
+
+    private VisionCalibrationProfile create(String key, String type, String name, String robotId,
+                                              String armDeviceId, String cameraDeviceId, String toolCode,
+                                              CalibrationResult result, JSONObject metadata, String operator) {
+        VisionCalibrationProfile latest = baseMapper.selectOne(new LambdaQueryWrapper<VisionCalibrationProfile>()
+                .eq(VisionCalibrationProfile::getProfileKey, key)
+                .orderByDesc(VisionCalibrationProfile::getVersionNo).last("limit 1"));
+        VisionCalibrationProfile profile = new VisionCalibrationProfile();
+        profile.setProfileType(type);
+        profile.setProfileKey(key);
+        profile.setProfileName(StrUtil.blankToDefault(name, HAND_EYE.equals(type) ? "手眼标定" : toolCode));
+        profile.setVersionNo(latest == null ? 1 : latest.getVersionNo() + 1);
+        profile.setRobotId(robotId);
+        profile.setArmDeviceId(armDeviceId);
+        profile.setCameraDeviceId(cameraDeviceId);
+        profile.setToolCode(toolCode);
+        profile.setRawMatrixJson(VisionCalibrationMath.toJson(result.matrix()));
+        profile.setCorrectionJson(JSON.toJSONString(new VisionCalibrationRequests.Correction()));
+        profile.setEffectiveMatrixJson(profile.getRawMatrixJson());
+        profile.setMeanErrorMm(decimal(result.meanErrorMm()));
+        profile.setMaxErrorMm(decimal(result.maxErrorMm()));
+        profile.setSampleCount(result.sampleCount());
+        profile.setMetadataJson(metadata == null ? null : metadata.toJSONString());
+        profile.setForcedEnabled(false);
+        profile.setStatus(result.maxErrorMm() <= 1.0 ? ACTIVE : DRAFT);
+        profile.setVerifiedTime(LocalDateTime.now());
+        profile.setCreateBy(operator);
+        if (ACTIVE.equals(profile.getStatus())) deactivateOthers(key, null, operator);
+        baseMapper.insert(profile);
+        return profile;
+    }
+
+    @Transactional(rollbackFor = Exception.class)
+    public VisionCalibrationProfile updateCorrection(String id, VisionCalibrationRequests.Correction correction,
+                                                      String operator) {
+        VisionCalibrationProfile profile = require(id);
+        double[][] effective = VisionCalibrationMath.applyCorrection(
+                VisionCalibrationMath.parseMatrix(profile.getRawMatrixJson()), correction);
+        profile.setCorrectionJson(JSON.toJSONString(correction));
+        profile.setEffectiveMatrixJson(VisionCalibrationMath.toJson(effective));
+        profile.setStatus(DRAFT);
+        profile.setForcedEnabled(false);
+        profile.setVerifiedTime(null);
+        profile.setUpdateBy(operator);
+        baseMapper.updateById(profile);
+        return profile;
+    }
+
+    @Transactional(rollbackFor = Exception.class)
+    public VisionCalibrationProfile activate(String id, boolean forced, String operator) {
+        VisionCalibrationProfile profile = require(id);
+        if (profile.getVerifiedTime() == null) throw new GlobalException("人工修正后必须先完成快速验证");
+        if (!forced && profile.getMaxErrorMm() != null && profile.getMaxErrorMm().doubleValue() > 1.0) {
+            throw new GlobalException("验证误差超过1毫米，请人工确认后强制启用");
+        }
+        deactivateOthers(profile.getProfileKey(), profile.getId(), operator);
+        profile.setStatus(ACTIVE);
+        profile.setForcedEnabled(forced);
+        profile.setUpdateBy(operator);
+        baseMapper.updateById(profile);
+        return profile;
+    }
+
+    public VisionCalibrationProfile markVerified(String id, double meanErrorMm, double maxErrorMm,
+                                                  int sampleCount, String operator) {
+        VisionCalibrationProfile profile = require(id);
+        profile.setMeanErrorMm(decimal(meanErrorMm));
+        profile.setMaxErrorMm(decimal(maxErrorMm));
+        profile.setSampleCount(sampleCount);
+        profile.setVerifiedTime(LocalDateTime.now());
+        profile.setUpdateBy(operator);
+        baseMapper.updateById(profile);
+        return profile;
+    }
+
+    @Transactional(rollbackFor = Exception.class)
+    public void invalidate(String id, String operator) {
+        VisionCalibrationProfile profile = require(id);
+        profile.setStatus(INVALID);
+        profile.setUpdateBy(operator);
+        baseMapper.updateById(profile);
+    }
+
+    private void deactivateOthers(String key, String exceptId, String operator) {
+        LambdaUpdateWrapper<VisionCalibrationProfile> update = new LambdaUpdateWrapper<VisionCalibrationProfile>()
+                .eq(VisionCalibrationProfile::getProfileKey, key)
+                .eq(VisionCalibrationProfile::getStatus, ACTIVE)
+                .set(VisionCalibrationProfile::getStatus, DRAFT)
+                .set(VisionCalibrationProfile::getUpdateBy, operator)
+                .set(VisionCalibrationProfile::getUpdateTime, LocalDateTime.now());
+        if (StrUtil.isNotBlank(exceptId)) update.ne(VisionCalibrationProfile::getId, exceptId);
+        baseMapper.update(null, update);
+    }
+
+    private String profileKey(String type, String... parts) {
+        return DigestUtil.sha256Hex(type + "|" + String.join("|", parts));
+    }
+
+    private BigDecimal decimal(double value) {
+        return BigDecimal.valueOf(value).setScale(4, RoundingMode.HALF_UP);
+    }
+}
diff --git a/cmvr-iot-test/src/main/java/com/cmvr/test/vision/service/VisionCalibrationRecordService.java b/cmvr-iot-test/src/main/java/com/cmvr/test/vision/service/VisionCalibrationRecordService.java
new file mode 100644
index 0000000..1955c76
--- /dev/null
+++ b/cmvr-iot-test/src/main/java/com/cmvr/test/vision/service/VisionCalibrationRecordService.java
@@ -0,0 +1,11 @@
+package com.cmvr.test.vision.service;
+
+import com.baomidou.mybatisplus.spring.service.impl.ServiceImpl;
+import com.cmvr.test.vision.domain.VisionCalibrationRecord;
+import com.cmvr.test.vision.mapper.VisionCalibrationRecordMapper;
+import org.springframework.stereotype.Service;
+
+@Service
+public class VisionCalibrationRecordService
+        extends ServiceImpl<VisionCalibrationRecordMapper, VisionCalibrationRecord> {
+}
diff --git a/cmvr-iot-test/src/main/java/com/cmvr/test/vision/service/VisionCalibrationSessionService.java b/cmvr-iot-test/src/main/java/com/cmvr/test/vision/service/VisionCalibrationSessionService.java
new file mode 100644
index 0000000..5c60114
--- /dev/null
+++ b/cmvr-iot-test/src/main/java/com/cmvr/test/vision/service/VisionCalibrationSessionService.java
@@ -0,0 +1,563 @@
+package com.cmvr.test.vision.service;
+
+import cn.hutool.core.util.StrUtil;
+import com.alibaba.fastjson2.JSON;
+import com.alibaba.fastjson2.JSONArray;
+import com.alibaba.fastjson2.JSONObject;
+import com.cmvr.common.exception.GlobalException;
+import com.cmvr.edge.client.model.EdgeCommonVO;
+import com.cmvr.edge.client.model.camera.EdgeCameraRgbdSnapshot;
+import com.cmvr.edge.client.service.EdgeArmService;
+import com.cmvr.edge.client.service.EdgeCameraService;
+import com.cmvr.test.vision.domain.VisionCalibrationProfile;
+import com.cmvr.test.vision.domain.VisionCalibrationRecord;
+import com.cmvr.test.vision.marker.MarkerObservation;
+import com.cmvr.test.vision.marker.SquareMarkerDetector;
+import com.cmvr.test.vision.math.CalibrationResult;
+import com.cmvr.test.vision.math.CalibrationSample;
+import com.cmvr.test.vision.math.VisionCalibrationMath;
+import com.cmvr.test.vision.model.VisionCalibrationRequests;
+import com.cmvr.test.vision.model.VisionCalibrationSessionView;
+import jakarta.annotation.PreDestroy;
+import lombok.RequiredArgsConstructor;
+import lombok.extern.slf4j.Slf4j;
+import org.springframework.stereotype.Service;
+
+import java.time.LocalDateTime;
+import java.util.ArrayList;
+import java.util.List;
+import java.util.Map;
+import java.util.UUID;
+import java.util.concurrent.ConcurrentHashMap;
+import java.util.concurrent.ExecutorService;
+import java.util.concurrent.Executors;
+import java.util.concurrent.ScheduledExecutorService;
+import java.util.concurrent.TimeUnit;
+import java.util.concurrent.atomic.AtomicBoolean;
+
+@Service
+@RequiredArgsConstructor
+@Slf4j
+public class VisionCalibrationSessionService {
+    private static final long HEARTBEAT_TIMEOUT_MS = 6_000;
+    private static final int HAND_EYE_SAMPLE_COUNT = 12;
+    private final EdgeArmService armService;
+    private final EdgeCameraService cameraService;
+    private final SquareMarkerDetector markerDetector;
+    private final VisionCalibrationProfileService profileService;
+    private final VisionCalibrationRecordService recordService;
+    private final Map<String, Session> sessions = new ConcurrentHashMap<>();
+    private final ExecutorService worker = Executors.newCachedThreadPool(runnable -> {
+        Thread thread = new Thread(runnable, "vision-calibration-worker");
+        thread.setDaemon(true);
+        return thread;
+    });
+    private final ScheduledExecutorService watchdog = Executors.newSingleThreadScheduledExecutor(runnable -> {
+        Thread thread = new Thread(runnable, "vision-calibration-watchdog");
+        thread.setDaemon(true);
+        return thread;
+    });
+
+    public VisionCalibrationSessionView startHandEye(VisionCalibrationRequests.HandEyeStart request, String operator) {
+        validateHandEye(request);
+        Session session = create("HAND_EYE", request.getRobotId(), request.getArmDeviceId(),
+                request.getCameraDeviceId(), null, operator, HAND_EYE_SAMPLE_COUNT);
+        session.handEyeRequest = request;
+        sessions.put(session.id, session);
+        worker.submit(() -> runHandEye(session));
+        return view(session);
+    }
+
+    public VisionCalibrationSessionView startTool(VisionCalibrationRequests.ToolStart request, String operator) {
+        if (StrUtil.hasBlank(request.getToolCode(), request.getRobotId(), request.getArmDeviceId(), request.getCameraDeviceId())) {
+            throw new GlobalException("工具名称、机器人、机械臂和相机不能为空");
+        }
+        VisionCalibrationProfile handEye = profileService.activeHandEye(
+                request.getRobotId(), request.getArmDeviceId(), request.getCameraDeviceId());
+        Session session = create("TOOL_TCP", request.getRobotId(), request.getArmDeviceId(),
+                request.getCameraDeviceId(), request.getToolCode(), operator, 6);
+        session.toolRequest = request;
+        session.handEyeProfile = handEye;
+        try {
+            session.message = "正在识别工具方向标定码";
+            double[][] flange = currentFlange(session);
+            MarkerObservation marker = captureMarker(session);
+            double[][] baseToCamera = VisionCalibrationMath.multiply(flange,
+                    VisionCalibrationMath.parseMatrix(handEye.getEffectiveMatrixJson()));
+            session.planeNormalBase = rotate(baseToCamera, marker.normalCamera());
+            session.planeXBase = rotate(baseToCamera, marker.xAxisCamera());
+            session.imageUrl = marker.imageUrl();
+            session.markerCorners = marker.pixelCorners();
+            session.imageWidth = marker.width(); session.imageHeight = marker.height();
+            session.message = "标定平面已记录，请将工具中心从不同角度对准页面标示的规范角点";
+            updateRecord(session);
+        } catch (RuntimeException exception) {
+            fail(session, exception);
+        } finally {
+            sessions.put(session.id, session);
+        }
+        return view(session);
+    }
+
+    public VisionCalibrationSessionView startVerification(String profileId, String operator) {
+        VisionCalibrationProfile profile = profileService.require(profileId);
+        if (VisionCalibrationProfileService.HAND_EYE.equals(profile.getProfileType())) {
+            Session session = create("HAND_EYE_VERIFY", profile.getRobotId(), profile.getArmDeviceId(),
+                    profile.getCameraDeviceId(), null, operator, 4);
+            session.profileToVerify = profile;
+            VisionCalibrationRequests.HandEyeStart request = new VisionCalibrationRequests.HandEyeStart();
+            request.setRobotId(profile.getRobotId()); request.setArmDeviceId(profile.getArmDeviceId());
+            request.setCameraDeviceId(profile.getCameraDeviceId()); request.setTranslationRangeMm(15.0);
+            request.setRotationRangeDeg(6.0); request.setVelocity(0.025); request.setAcceleration(0.06);
+            request.setSettleTimeMs(500);
+            session.handEyeRequest = request;
+            sessions.put(session.id, session);
+            worker.submit(() -> runHandEyeVerification(session));
+            return view(session);
+        }
+        Session session = create("TOOL_VERIFY", profile.getRobotId(), profile.getArmDeviceId(),
+                profile.getCameraDeviceId(), profile.getToolCode(), operator, 4);
+        session.profileToVerify = profile;
+        session.handEyeProfile = profileService.activeHandEye(
+                profile.getRobotId(), profile.getArmDeviceId(), profile.getCameraDeviceId());
+        try {
+            captureToolPlane(session);
+            session.message = "请将工具中心从3个不同角度对准规范角点并依次记录";
+        } finally {
+            sessions.put(session.id, session);
+        }
+        return view(session);
+    }
+
+    public VisionCalibrationSessionView recordToolPivot(String sessionId) {
+        Session session = requireRunningTool(sessionId);
+        double[][] pose = currentFlange(session);
+        if (session.flangeSamples.stream().anyMatch(previous -> rotationDistanceDeg(previous, pose) < 4.0)) {
+            throw new GlobalException("本次末端角度与已有样本过于接近，请明显改变姿态后再记录");
+        }
+        session.flangeSamples.add(pose);
+        session.currentStep = session.flangeSamples.size();
+        int required = "TOOL_VERIFY".equals(session.type) ? 3 : 5;
+        session.message = session.flangeSamples.size() < required
+                ? "已记录姿态 " + session.flangeSamples.size() + "/" + required + "，请改变末端角度并继续对准同一角点"
+                : "定点姿态已足够，请将工具垂直贴合标定码平面并记录安全方向";
+        updateRecord(session);
+        return view(session);
+    }
+
+    public VisionCalibrationSessionView recordToolDirectionAndFinish(String sessionId) {
+        Session session = requireRunningTool(sessionId);
+        int required = "TOOL_VERIFY".equals(session.type) ? 3 : 5;
+        if (session.flangeSamples.size() < required) throw new GlobalException("请先记录至少" + required + "个定点旋转姿态");
+        double[][] directionPose = currentFlange(session);
+        if ("TOOL_VERIFY".equals(session.type)) {
+            CalibrationResult verification = verifyTool(session, directionPose);
+            profileService.markVerified(session.profileToVerify.getId(), verification.meanErrorMm(),
+                    verification.maxErrorMm(), session.flangeSamples.size(), session.operator);
+            complete(session, session.profileToVerify, verification);
+            session.message = session.directionErrorDeg <= 1.0 && verification.maxErrorMm() <= 1.0
+                    ? "工具TCP验证通过，可以启用" : "验证完成但误差超限，可继续微调或人工强制启用";
+            return view(session);
+        }
+        CalibrationResult pivot = VisionCalibrationMath.solveToolPivot(session.flangeSamples);
+        double[][] matrix = VisionCalibrationMath.applyToolDirection(
+                pivot.matrix(), directionPose, session.planeNormalBase, session.planeXBase);
+        JSONObject metadata = new JSONObject()
+                .fluentPut("handEyeProfileId", session.handEyeProfile.getId())
+                .fluentPut("handEyeVersion", session.handEyeProfile.getVersionNo())
+                .fluentPut("markerImageUrl", session.imageUrl)
+                .fluentPut("safeDirectionPose", matrixJson(directionPose))
+                .fluentPut("planeNormalBase", session.planeNormalBase)
+                .fluentPut("planeXBase", session.planeXBase);
+        VisionCalibrationProfile profile = profileService.createTool(
+                session.toolRequest, pivot, matrix, metadata, session.operator);
+        complete(session, profile, pivot);
+        return view(session);
+    }
+
+    private void captureToolPlane(Session session) {
+        try {
+            double[][] flange = currentFlange(session);
+            MarkerObservation marker = captureMarker(session);
+            double[][] baseToCamera = VisionCalibrationMath.multiply(flange,
+                    VisionCalibrationMath.parseMatrix(session.handEyeProfile.getEffectiveMatrixJson()));
+            session.planeNormalBase = rotate(baseToCamera, marker.normalCamera());
+            session.planeXBase = rotate(baseToCamera, marker.xAxisCamera());
+            session.imageUrl = marker.imageUrl(); session.markerCorners = marker.pixelCorners();
+            session.imageWidth = marker.width(); session.imageHeight = marker.height();
+            updateRecord(session);
+        } catch (RuntimeException exception) {
+            fail(session, exception);
+            throw exception;
+        }
+    }
+
+    private CalibrationResult verifyTool(Session session, double[][] directionPose) {
+        double[][] toolMatrix = VisionCalibrationMath.parseMatrix(session.profileToVerify.getEffectiveMatrixJson());
+        double[] tcp = {toolMatrix[0][3], toolMatrix[1][3], toolMatrix[2][3]};
+        List<double[]> contactPoints = session.flangeSamples.stream()
+                .map(pose -> VisionCalibrationMath.transform(pose, tcp)).toList();
+        double[] center = new double[3];
+        contactPoints.forEach(point -> { for (int axis = 0; axis < 3; axis++) center[axis] += point[axis] / contactPoints.size(); });
+        List<Double> errors = contactPoints.stream().map(point -> distance(point, center) * 1000.0).toList();
+        double[] toolZFlange = {toolMatrix[0][2], toolMatrix[1][2], toolMatrix[2][2]};
+        double[] toolZBase = normalize(rotate(directionPose, toolZFlange));
+        double[] expected = normalize(scale(session.planeNormalBase, -1));
+        session.directionErrorDeg = Math.toDegrees(Math.acos(Math.max(-1, Math.min(1, dot(toolZBase, expected)))));
+        return new CalibrationResult(toolMatrix, errors.stream().mapToDouble(Double::doubleValue).average().orElse(0),
+                errors.stream().mapToDouble(Double::doubleValue).max().orElse(0), contactPoints.size(), center);
+    }
+
+    public VisionCalibrationSessionView heartbeat(String sessionId) {
+        Session session = require(sessionId);
+        session.lastHeartbeat = System.currentTimeMillis();
+        return view(session);
+    }
+
+    public VisionCalibrationSessionView status(String sessionId) {
+        return view(require(sessionId));
+    }
+
+    public VisionCalibrationSessionView stop(String sessionId) {
+        Session session = require(sessionId);
+        cancel(session, "用户停止标定", true);
+        return view(session);
+    }
+
+    private void runHandEye(Session session) {
+        double[][] initialPose = null;
+        try {
+            session.message = "正在记录初始位姿并识别标定码";
+            initialPose = currentFlange(session);
+            session.initialPose = initialPose;
+            List<double[][]> targets = calibrationTargets(initialPose, session.handEyeRequest);
+            for (int index = 0; index < targets.size(); index++) {
+                checkRunning(session);
+                if (index > 0) {
+                    session.message = "正在移动到标定姿态 " + (index + 1) + "/" + targets.size();
+                    move(session, targets.get(index), session.handEyeRequest.getVelocity(),
+                            session.handEyeRequest.getAcceleration());
+                    sleep(session.handEyeRequest.getSettleTimeMs());
+                }
+                checkRunning(session);
+                double[][] flange = currentFlange(session);
+                MarkerObservation marker = captureMarker(session);
+                session.handEyeSamples.add(new CalibrationSample(flange, marker.centerCamera()));
+                session.imageUrl = marker.imageUrl();
+                session.markerCorners = marker.pixelCorners();
+                session.imageWidth = marker.width(); session.imageHeight = marker.height();
+                session.currentStep = session.handEyeSamples.size();
+                session.message = "已采集 " + session.currentStep + "/" + targets.size()
+                        + "，平面拟合误差 " + String.format("%.2f", marker.planeRmseMm()) + " mm";
+                updateRecord(session);
+            }
+            CalibrationResult solved = solveWithValidation(session.handEyeSamples);
+            JSONObject metadata = new JSONObject()
+                    .fluentPut("markerSizeMm", session.handEyeRequest.getMarkerSizeMm())
+                    .fluentPut("translationRangeMm", session.handEyeRequest.getTranslationRangeMm())
+                    .fluentPut("rotationRangeDeg", session.handEyeRequest.getRotationRangeDeg())
+                    .fluentPut("fixedMarkerPoint", solved.fixedPoint())
+                    .fluentPut("validationSamples", 3);
+            session.message = "标定完成，正在返回初始位姿";
+            move(session, initialPose, session.handEyeRequest.getVelocity(), session.handEyeRequest.getAcceleration());
+            VisionCalibrationProfile profile = profileService.createHandEye(
+                    session.handEyeRequest, solved, metadata, session.operator);
+            complete(session, profile, solved);
+        } catch (CancelledException ignored) {
+            // The watchdog or stop endpoint already recorded the terminal state.
+        } catch (RuntimeException exception) {
+            fail(session, exception);
+        }
+    }
+
+    private void runHandEyeVerification(Session session) {
+        try {
+            double[][] initial = currentFlange(session);
+            session.initialPose = initial;
+            List<double[][]> targets = calibrationTargets(initial, session.handEyeRequest).subList(0, 4);
+            double[][] cameraTransform = VisionCalibrationMath.parseMatrix(session.profileToVerify.getEffectiveMatrixJson());
+            List<double[]> markerPoints = new ArrayList<>();
+            for (int index = 0; index < targets.size(); index++) {
+                checkRunning(session);
+                if (index > 0) {
+                    session.message = "正在执行验证姿态 " + (index + 1) + "/4";
+                    move(session, targets.get(index), session.handEyeRequest.getVelocity(), session.handEyeRequest.getAcceleration());
+                    sleep(session.handEyeRequest.getSettleTimeMs());
+                }
+                double[][] flange = currentFlange(session);
+                MarkerObservation marker = captureMarker(session);
+                markerPoints.add(VisionCalibrationMath.transform(
+                        VisionCalibrationMath.multiply(flange, cameraTransform), marker.centerCamera()));
+                session.currentStep = index + 1; session.imageUrl = marker.imageUrl();
+                session.markerCorners = marker.pixelCorners(); updateRecord(session);
+                session.imageWidth = marker.width(); session.imageHeight = marker.height();
+            }
+            double[] center = new double[3];
+            markerPoints.forEach(point -> { for (int axis = 0; axis < 3; axis++) center[axis] += point[axis] / markerPoints.size(); });
+            List<Double> errors = markerPoints.stream().map(point -> distance(point, center) * 1000.0).toList();
+            CalibrationResult result = new CalibrationResult(cameraTransform,
+                    errors.stream().mapToDouble(Double::doubleValue).average().orElse(0),
+                    errors.stream().mapToDouble(Double::doubleValue).max().orElse(0), markerPoints.size(), center);
+            move(session, initial, session.handEyeRequest.getVelocity(), session.handEyeRequest.getAcceleration());
+            profileService.markVerified(session.profileToVerify.getId(), result.meanErrorMm(), result.maxErrorMm(),
+                    result.sampleCount(), session.operator);
+            complete(session, session.profileToVerify, result);
+            session.message = result.maxErrorMm() <= 1.0 ? "快速验证通过，可以启用" : "验证完成但误差超过1毫米";
+        } catch (CancelledException ignored) {
+        } catch (RuntimeException exception) {
+            fail(session, exception);
+        }
+    }
+
+    private CalibrationResult solveWithValidation(List<CalibrationSample> samples) {
+        List<CalibrationSample> training = samples.subList(0, samples.size() - 3);
+        List<CalibrationSample> validation = samples.subList(samples.size() - 3, samples.size());
+        CalibrationResult solved = VisionCalibrationMath.solveHandEye(training, null);
+        List<Double> errors = new ArrayList<>();
+        for (CalibrationSample sample : validation) {
+            double[] predicted = VisionCalibrationMath.transform(
+                    VisionCalibrationMath.multiply(sample.baseToFlange(), solved.matrix()), sample.cameraPoint());
+            errors.add(distance(predicted, solved.fixedPoint()) * 1000.0);
+        }
+        double mean = errors.stream().mapToDouble(Double::doubleValue).average().orElse(0);
+        double max = errors.stream().mapToDouble(Double::doubleValue).max().orElse(0);
+        return new CalibrationResult(solved.matrix(), mean, max, samples.size(), solved.fixedPoint());
+    }
+
+    private List<double[][]> calibrationTargets(double[][] initial, VisionCalibrationRequests.HandEyeStart request) {
+        double t = clamp(request.getTranslationRangeMm(), 5, 20) / 1000.0;
+        double r = Math.toRadians(clamp(request.getRotationRangeDeg(), 3, 10));
+        double[][] deltas = {
+                {0,0,0,0,0,0}, {t,0,0,r,0,0}, {-t,0,0,-r,0,0},
+                {0,t,0,0,r,0}, {0,-t,0,0,-r,0}, {0,0,t,0,0,r},
+                {0,0,-t,0,0,-r}, {t,t,0,r,-r,0}, {-t,t,0,-r,-r,0},
+                {t,-t,0,r,r,0}, {-t,-t,0,-r,r,0}, {0,t,t,r,0,r}
+        };
+        List<double[][]> result = new ArrayList<>();
+        for (double[] delta : deltas) {
+            double[][] transform = VisionCalibrationMath.fromEuler(delta[3], delta[4], delta[5]);
+            transform[0][3] = delta[0]; transform[1][3] = delta[1]; transform[2][3] = delta[2];
+            result.add(VisionCalibrationMath.multiply(initial, transform));
+        }
+        return result;
+    }
+
+    private MarkerObservation captureMarker(Session session) {
+        EdgeCameraRgbdSnapshot snapshot = cameraService.captureRgbd(target(session.robotId, session.cameraDeviceId));
+        return markerDetector.detect(snapshot);
+    }
+
+    private double[][] currentFlange(Session session) {
+        return VisionCalibrationMath.fromCartesianPose(armService.getPose(
+                target(session.robotId, session.armDeviceId), null, null).getPose());
+    }
+
+    private void move(Session session, double[][] target, Double velocity, Double acceleration) {
+        checkRunning(session);
+        double[] pose = VisionCalibrationMath.pose(target);
+        // Aubo consumes ZYX Euler angles. Unwrap each target angle around the
+        // current angle so crossing +/-PI never turns a small move into a full turn.
+        double[] current = VisionCalibrationMath.pose(currentFlange(session));
+        for (int index = 3; index < 6; index++) {
+            pose[index] = unwrapAngle(pose[index], current[index]);
+        }
+        armService.moveL(target(session.robotId, session.armDeviceId), pose[0], pose[1], pose[2],
+                pose[3], pose[4], pose[5], "BASE", clamp(velocity, 0.003, 0.03),
+                clamp(acceleration, 0.01, 0.08), 0.0);
+    }
+
+    private double unwrapAngle(double target, double reference) {
+        double twoPi = Math.PI * 2.0;
+        while (target - reference > Math.PI) target -= twoPi;
+        while (target - reference < -Math.PI) target += twoPi;
+        return target;
+    }
+
+    private Session create(String type, String robotId, String armDeviceId, String cameraDeviceId,
+                           String toolCode, String operator, int totalSteps) {
+        Session session = new Session();
+        session.id = UUID.randomUUID().toString().replace("-", "");
+        session.type = type; session.robotId = robotId; session.armDeviceId = armDeviceId;
+        session.cameraDeviceId = cameraDeviceId; session.toolCode = toolCode; session.operator = operator;
+        session.totalSteps = totalSteps; session.lastHeartbeat = System.currentTimeMillis();
+        VisionCalibrationRecord record = new VisionCalibrationRecord();
+        record.setId(session.id); record.setSessionType(type); record.setRobotId(robotId);
+        record.setArmDeviceId(armDeviceId); record.setCameraDeviceId(cameraDeviceId); record.setToolCode(toolCode);
+        record.setStatus("RUNNING"); record.setProgressStep(0); record.setTotalSteps(totalSteps);
+        record.setStartedTime(LocalDateTime.now()); record.setCreateBy(operator);
+        recordService.save(record);
+        return session;
+    }
+
+    private void complete(Session session, VisionCalibrationProfile profile, CalibrationResult result) {
+        session.running.set(false);
+        session.status = "COMPLETED"; session.profileId = profile.getId();
+        session.meanErrorMm = result.meanErrorMm(); session.maxErrorMm = result.maxErrorMm();
+        session.message = result.maxErrorMm() <= 1.0 ? "标定和验证通过，方案已自动启用" : "标定完成但误差超过1毫米，请人工微调或强制启用";
+        VisionCalibrationRecord record = recordService.getById(session.id);
+        record.setProfileId(profile.getId()); record.setStatus("COMPLETED");
+        record.setFinishedTime(LocalDateTime.now()); record.setProgressStep(session.currentStep);
+        record.setSamplesJson(samplesJson(session));
+        record.setResultJson(new JSONObject().fluentPut("profileId", profile.getId())
+                .fluentPut("meanErrorMm", result.meanErrorMm()).fluentPut("maxErrorMm", result.maxErrorMm())
+                .fluentPut("matrix", JSON.parseArray(profile.getEffectiveMatrixJson())).toJSONString());
+        recordService.updateById(record);
+    }
+
+    private void fail(Session session, RuntimeException exception) {
+        if (!session.running.compareAndSet(true, false)) return;
+        session.status = "FAILED"; session.message = exception.getMessage();
+        log.error("视觉标定会话失败: {}", session.id, exception);
+        if (session.type.startsWith("HAND_EYE")) {
+            try { armService.stopMotion(target(session.robotId, session.armDeviceId)); }
+            catch (RuntimeException stopException) { log.error("标定失败后停止机械臂失败: {}", session.id, stopException); }
+        }
+        VisionCalibrationRecord record = recordService.getById(session.id);
+        record.setStatus("FAILED"); record.setErrorMessage(StrUtil.maxLength(exception.getMessage(), 1000));
+        record.setFinishedTime(LocalDateTime.now()); record.setProgressStep(session.currentStep);
+        record.setSamplesJson(samplesJson(session)); recordService.updateById(record);
+    }
+
+    private void cancel(Session session, String reason, boolean stopArm) {
+        if (!session.running.compareAndSet(true, false)) return;
+        session.status = "CANCELLED"; session.message = reason;
+        if (stopArm) {
+            try { armService.stopMotion(target(session.robotId, session.armDeviceId)); }
+            catch (RuntimeException exception) { log.error("停止标定机械臂失败: {}", session.id, exception); }
+        }
+        VisionCalibrationRecord record = recordService.getById(session.id);
+        record.setStatus("CANCELLED"); record.setErrorMessage(reason); record.setFinishedTime(LocalDateTime.now());
+        record.setProgressStep(session.currentStep); record.setSamplesJson(samplesJson(session)); recordService.updateById(record);
+    }
+
+    private void updateRecord(Session session) {
+        VisionCalibrationRecord record = recordService.getById(session.id);
+        record.setProgressStep(session.currentStep); record.setSamplesJson(samplesJson(session));
+        recordService.updateById(record);
+    }
+
+    private String samplesJson(Session session) {
+        JSONArray values = new JSONArray();
+        session.handEyeSamples.forEach(sample -> values.add(new JSONObject()
+                .fluentPut("baseToFlange", matrixJson(sample.baseToFlange()))
+                .fluentPut("cameraPoint", sample.cameraPoint())));
+        session.flangeSamples.forEach(sample -> values.add(new JSONObject().fluentPut("baseToFlange", matrixJson(sample))));
+        return values.toJSONString();
+    }
+
+    private JSONArray matrixJson(double[][] matrix) { return JSON.parseArray(VisionCalibrationMath.toJson(matrix)); }
+
+    private Session require(String id) {
+        Session session = sessions.get(id);
+        if (session == null) throw new GlobalException("标定会话不存在或已过期");
+        return session;
+    }
+
+    private Session requireRunning(String id, String type) {
+        Session session = require(id);
+        if (!type.equals(session.type)) throw new GlobalException("标定会话类型不匹配");
+        checkRunning(session);
+        return session;
+    }
+
+    private Session requireRunningTool(String id) {
+        Session session = require(id);
+        if (!"TOOL_TCP".equals(session.type) && !"TOOL_VERIFY".equals(session.type))
+            throw new GlobalException("标定会话类型不匹配");
+        checkRunning(session);
+        return session;
+    }
+
+    private void checkRunning(Session session) {
+        if (!session.running.get()) throw new CancelledException();
+    }
+
+    private void validateHandEye(VisionCalibrationRequests.HandEyeStart request) {
+        if (StrUtil.hasBlank(request.getRobotId(), request.getArmDeviceId(), request.getCameraDeviceId()))
+            throw new GlobalException("机器人、机械臂和相机不能为空");
+        request.setTranslationRangeMm(clamp(request.getTranslationRangeMm(), 5, 20));
+        request.setRotationRangeDeg(clamp(request.getRotationRangeDeg(), 3, 10));
+        request.setVelocity(clamp(request.getVelocity(), 0.003, 0.03));
+        request.setAcceleration(clamp(request.getAcceleration(), 0.01, 0.08));
+        request.setSettleTimeMs(Math.max(200, Math.min(3000, request.getSettleTimeMs() == null ? 500 : request.getSettleTimeMs())));
+    }
+
+    private EdgeCommonVO target(String robotId, String deviceId) {
+        EdgeCommonVO value = new EdgeCommonVO(); value.setRobotId(robotId); value.setDeviceId(deviceId); return value;
+    }
+
+    private VisionCalibrationSessionView view(Session session) {
+        return VisionCalibrationSessionView.builder().sessionId(session.id).sessionType(session.type)
+                .status(session.status).currentStep(session.currentStep).totalSteps(session.totalSteps)
+                .sampleCount(session.handEyeSamples.size() + session.flangeSamples.size())
+                .message(session.message).imageUrl(session.imageUrl).profileId(session.profileId)
+                .meanErrorMm(session.meanErrorMm).maxErrorMm(session.maxErrorMm)
+                .directionErrorDeg(session.directionErrorDeg)
+                .markerCorners(session.markerCorners).imageWidth(session.imageWidth).imageHeight(session.imageHeight).build();
+    }
+
+    private double[] rotate(double[][] transform, double[] value) {
+        return new double[]{transform[0][0]*value[0]+transform[0][1]*value[1]+transform[0][2]*value[2],
+                transform[1][0]*value[0]+transform[1][1]*value[1]+transform[1][2]*value[2],
+                transform[2][0]*value[0]+transform[2][1]*value[1]+transform[2][2]*value[2]};
+    }
+
+    private double rotationDistanceDeg(double[][] left, double[][] right) {
+        double trace = 0;
+        for (int row = 0; row < 3; row++) for (int column = 0; column < 3; column++) trace += left[column][row] * right[column][row];
+        return Math.toDegrees(Math.acos(Math.max(-1, Math.min(1, (trace - 1) / 2))));
+    }
+
+    private double distance(double[] left, double[] right) {
+        double sum = 0; for (int index = 0; index < 3; index++) sum += Math.pow(left[index] - right[index], 2); return Math.sqrt(sum);
+    }
+    private double[] scale(double[] value, double factor) { return new double[]{value[0]*factor, value[1]*factor, value[2]*factor}; }
+    private double dot(double[] left, double[] right) { return left[0]*right[0]+left[1]*right[1]+left[2]*right[2]; }
+    private double[] normalize(double[] value) { double size = Math.sqrt(dot(value, value)); if (size < 1e-10) throw new GlobalException("工具方向无效"); return scale(value, 1/size); }
+
+    private void sleep(Integer milliseconds) {
+        try { Thread.sleep(milliseconds == null ? 500 : milliseconds); }
+        catch (InterruptedException exception) { Thread.currentThread().interrupt(); throw new CancelledException(); }
+    }
+
+    private double clamp(Double value, double min, double max) { return Math.max(min, Math.min(max, value == null ? min : value)); }
+
+    private void inspectHeartbeats() {
+        long now = System.currentTimeMillis();
+        sessions.values().stream().filter(session -> "RUNNING".equals(session.status))
+                .filter(session -> now - session.lastHeartbeat > HEARTBEAT_TIMEOUT_MS)
+                .forEach(session -> cancel(session, "页面连接中断，标定已自动停止", true));
+    }
+
+    {
+        watchdog.scheduleAtFixedRate(this::inspectHeartbeats, 2, 2, TimeUnit.SECONDS);
+    }
+
+    @PreDestroy
+    public void close() {
+        sessions.values().forEach(session -> cancel(session, "服务停止，标定已取消", true));
+        worker.shutdownNow(); watchdog.shutdownNow();
+    }
+
+    private static final class Session {
+        private String id, type, status = "RUNNING", message = "标定会话已创建", imageUrl, profileId;
+        private String robotId, armDeviceId, cameraDeviceId, toolCode, operator;
+        private int currentStep, totalSteps;
+        private long lastHeartbeat;
+        private Double meanErrorMm, maxErrorMm, directionErrorDeg;
+        private List<double[]> markerCorners = List.of();
+        private Integer imageWidth, imageHeight;
+        private final AtomicBoolean running = new AtomicBoolean(true);
+        private final List<CalibrationSample> handEyeSamples = new ArrayList<>();
+        private final List<double[][]> flangeSamples = new ArrayList<>();
+        private VisionCalibrationRequests.HandEyeStart handEyeRequest;
+        private VisionCalibrationRequests.ToolStart toolRequest;
+        private VisionCalibrationProfile handEyeProfile;
+        private VisionCalibrationProfile profileToVerify;
+        private double[][] initialPose;
+        private double[] planeNormalBase, planeXBase;
+    }
+
+    private static final class CancelledException extends RuntimeException {
+    }
+}
diff --git a/pom.xml b/pom.xml
index 45cb90f..8038de0 100644
--- a/pom.xml
+++ b/pom.xml
@@ -42,6 +42,7 @@
         <swagger.annotations.version>1.6.14</swagger.annotations.version>
         <netty.quic.version>4.2.16.Final</netty.quic.version>
         <nashorn.version>15.4</nashorn.version>
+        <opencv.version>4.9.0-0</opencv.version>
     </properties>
 
     <!-- 依赖声明 -->
@@ -217,6 +218,12 @@
                 <version>${cmvr-iot.version}</version>
             </dependency>
 
+            <dependency>
+                <groupId>org.openpnp</groupId>
+                <artifactId>opencv</artifactId>
+                <version>${opencv.version}</version>
+            </dependency>
+
             <dependency>
                 <groupId>com.cmvr</groupId>
                 <artifactId>cmvr-iot-device</artifactId>
diff --git a/sql/resource_center_menu.sql b/sql/resource_center_menu.sql
index d04a3f2..6a623e3 100644
--- a/sql/resource_center_menu.sql
+++ b/sql/resource_center_menu.sql
@@ -7,7 +7,8 @@ VALUES
  (2201, '机器人', 2200, 1, 'robot', 'resource/robot/index', '', 'ResourceRobot', 1, 0, 'C', '0', '0', 'resource:robot:list', 'monitor', 'admin', NOW(), '', NULL, '平台机器人身份与运行状态'),
  (2202, '设备', 2200, 2, 'device', 'resource/device/index', '', 'ResourceDevice', 1, 0, 'C', '0', '0', 'resource:device:list', 'list', 'admin', NOW(), '', NULL, '机器人设备发现与用途配置'),
  (2203, '地图', 2200, 3, 'map', 'resource/map/index', '', 'ResourceMap', 1, 0, 'C', '0', '0', 'resource:map:list', 'map', 'admin', NOW(), '', NULL, '仅有AGV能力的机器人需要绑定地图'),
- (2204, '边缘终端', 2200, 4, 'terminal', 'resource/terminal/index', '', 'ResourceTerminal', 1, 0, 'C', '0', '0', 'resource:terminal:list', 'server', 'admin', NOW(), '', NULL, 'QUIC边缘终端运行状态')
+ (2204, '边缘终端', 2200, 4, 'terminal', 'resource/terminal/index', '', 'ResourceTerminal', 1, 0, 'C', '0', '0', 'resource:terminal:list', 'server', 'admin', NOW(), '', NULL, 'QUIC边缘终端运行状态'),
+ (2205, '视觉标定管理', 2200, 5, 'vision-calibration', 'resource/visionCalibration/index', '', 'VisionCalibration', 1, 0, 'C', '0', '0', '', 'aim', 'admin', NOW(), '', NULL, '机器人手眼标定与工具TCP管理')
 ON DUPLICATE KEY UPDATE menu_name = VALUES(menu_name), parent_id = VALUES(parent_id), order_num = VALUES(order_num),
  path = VALUES(path), component = VALUES(component), route_name = VALUES(route_name), visible = VALUES(visible), status = VALUES(status),
  perms = VALUES(perms), icon = VALUES(icon), update_by = 'admin', update_time = NOW(), remark = VALUES(remark);
@@ -59,8 +60,11 @@ JOIN (
  UNION ALL SELECT 2204, 2241 UNION ALL SELECT 2204, 2242 UNION ALL SELECT 2204, 2243
 ) f ON f.parent_id = rm.menu_id;
 
-DELETE FROM sys_role_menu WHERE menu_id IN (2205, 2251, 2252, 2253);
-DELETE FROM sys_menu WHERE menu_id IN (2251, 2252, 2253, 2205);
+DELETE FROM sys_role_menu WHERE menu_id IN (2251, 2252, 2253);
+DELETE FROM sys_menu WHERE menu_id IN (2251, 2252, 2253);
+
+INSERT IGNORE INTO sys_role_menu(role_id, menu_id)
+SELECT role_id, 2205 FROM sys_role;
 
 -- Remove old device CRUD and inspection resource menus after permissions have been migrated.
 DELETE FROM sys_role_menu
diff --git a/sql/vision_calibration.sql b/sql/vision_calibration.sql
new file mode 100644
index 0000000..3d3c707
--- /dev/null
+++ b/sql/vision_calibration.sql
@@ -0,0 +1,73 @@
+-- Visual hand-eye and tool TCP calibration profiles. Apply once before using the management page.
+CREATE TABLE IF NOT EXISTS `te_vision_calibration_profile` (
+  `id` varchar(32) NOT NULL,
+  `profile_type` varchar(20) NOT NULL COMMENT 'HAND_EYE or TOOL_TCP',
+  `profile_key` varchar(100) NOT NULL COMMENT 'Stable logical key used by workflows',
+  `profile_name` varchar(100) NOT NULL,
+  `version_no` int NOT NULL DEFAULT 1,
+  `robot_id` varchar(128) NOT NULL,
+  `arm_device_id` varchar(128) NOT NULL,
+  `camera_device_id` varchar(128) DEFAULT NULL,
+  `tool_code` varchar(100) DEFAULT NULL,
+  `status` varchar(20) NOT NULL DEFAULT 'DRAFT' COMMENT 'DRAFT, ACTIVE, INVALID',
+  `forced_enabled` tinyint(1) NOT NULL DEFAULT 0,
+  `raw_matrix_json` text NOT NULL,
+  `correction_json` varchar(500) DEFAULT NULL COMMENT 'x/y/z mm and rx/ry/rz degree correction',
+  `effective_matrix_json` text NOT NULL,
+  `mean_error_mm` decimal(10,4) DEFAULT NULL,
+  `max_error_mm` decimal(10,4) DEFAULT NULL,
+  `sample_count` int NOT NULL DEFAULT 0,
+  `metadata_json` text DEFAULT NULL,
+  `verified_time` datetime DEFAULT NULL,
+  `create_by` varchar(64) DEFAULT NULL,
+  `create_time` datetime DEFAULT NULL,
+  `update_by` varchar(64) DEFAULT NULL,
+  `update_time` datetime DEFAULT NULL,
+  `remark` varchar(500) DEFAULT NULL,
+  PRIMARY KEY (`id`),
+  UNIQUE KEY `uk_vision_profile_version` (`profile_key`, `version_no`),
+  KEY `idx_vision_profile_devices` (`profile_type`, `robot_id`, `arm_device_id`, `camera_device_id`),
+  KEY `idx_vision_profile_tool` (`profile_type`, `robot_id`, `arm_device_id`, `tool_code`),
+  KEY `idx_vision_profile_active` (`profile_key`, `status`)
+) ENGINE=InnoDB DEFAULT CHARSET=utf8mb4 COMMENT='Visual calibration profile versions';
+
+CREATE TABLE IF NOT EXISTS `te_vision_calibration_record` (
+  `id` varchar(32) NOT NULL,
+  `profile_id` varchar(32) DEFAULT NULL,
+  `session_type` varchar(30) NOT NULL COMMENT 'HAND_EYE, HAND_EYE_VERIFY, TOOL_TCP',
+  `robot_id` varchar(128) NOT NULL,
+  `arm_device_id` varchar(128) NOT NULL,
+  `camera_device_id` varchar(128) DEFAULT NULL,
+  `tool_code` varchar(100) DEFAULT NULL,
+  `status` varchar(20) NOT NULL COMMENT 'RUNNING, COMPLETED, FAILED, CANCELLED',
+  `progress_step` int NOT NULL DEFAULT 0,
+  `total_steps` int NOT NULL DEFAULT 0,
+  `samples_json` longtext DEFAULT NULL,
+  `result_json` longtext DEFAULT NULL,
+  `error_message` varchar(1000) DEFAULT NULL,
+  `started_time` datetime DEFAULT NULL,
+  `finished_time` datetime DEFAULT NULL,
+  `create_by` varchar(64) DEFAULT NULL,
+  `create_time` datetime DEFAULT NULL,
+  `update_by` varchar(64) DEFAULT NULL,
+  `update_time` datetime DEFAULT NULL,
+  `remark` varchar(500) DEFAULT NULL,
+  PRIMARY KEY (`id`),
+  KEY `idx_vision_record_profile` (`profile_id`),
+  KEY `idx_vision_record_devices` (`robot_id`, `arm_device_id`, `camera_device_id`),
+  KEY `idx_vision_record_status` (`status`, `started_time`)
+) ENGINE=InnoDB DEFAULT CHARSET=utf8mb4 COMMENT='Visual calibration execution and verification records';
+
+INSERT INTO sys_menu(menu_id, menu_name, parent_id, order_num, path, component, query, route_name,
+                     is_frame, is_cache, menu_type, visible, status, perms, icon,
+                     create_by, create_time, update_by, update_time, remark)
+VALUES
+ (2205, '视觉标定管理', 2200, 5, 'vision-calibration', 'resource/visionCalibration/index', '', 'VisionCalibration',
+  1, 0, 'C', '0', '0', '', 'aim', 'admin', NOW(), '', NULL, '机器人手眼标定与工具TCP管理')
+ON DUPLICATE KEY UPDATE menu_name=VALUES(menu_name), parent_id=VALUES(parent_id), order_num=VALUES(order_num),
+ path=VALUES(path), component=VALUES(component), route_name=VALUES(route_name), visible=VALUES(visible),
+ status=VALUES(status), perms=VALUES(perms), update_by='admin', update_time=NOW(), remark=VALUES(remark);
+
+-- Calibration management is available to every application role by product requirement.
+INSERT IGNORE INTO sys_role_menu(role_id, menu_id)
+SELECT role_id, 2205 FROM sys_role;