cmvr_ai_lab/scripts/train.py

"""Script to train RL agent with RSL-RL."""

"""Launch Isaac Sim Simulator first."""

import argparse
import sys
from pathlib import Path

from isaaclab.app import AppLauncher

# local imports
import cli_args  # isort: skip

# Prefer this checkout over any other editable ``engineai_lab`` installation.
REPO_ROOT = Path(__file__).resolve().parents[1]
SOURCE_ROOT = REPO_ROOT / "source"
if str(SOURCE_ROOT) not in sys.path:
    sys.path.insert(0, str(SOURCE_ROOT))

# add argparse arguments
parser = argparse.ArgumentParser(description="Train an RL agent with RSL-RL.")
parser.add_argument("--num_envs", type=int, default=None, help="Number of environments to simulate.")
parser.add_argument("--task", type=str, default=None, help="Name of the task.")
parser.add_argument("--seed", type=int, default=None, help="Seed used for the environment")
parser.add_argument("--max_iterations", type=int, default=None, help="RL Policy training iterations.")
parser.add_argument(
    "--distributed", action="store_true", default=False, help="Run training with multiple GPUs or nodes."
)

# append RSL-RL cli arguments
cli_args.add_rsl_rl_args(parser)
# append AppLauncher cli args
AppLauncher.add_app_launcher_args(parser)
args_cli, hydra_args = parser.parse_known_args()


# clear out sys.argv for Hydra
sys.argv = [sys.argv[0]] + hydra_args

# launch omniverse app
app_launcher = AppLauncher(args_cli)
simulation_app = app_launcher.app

"""Rest everything follows."""

import gymnasium as gym
import os
import torch
from datetime import datetime

from isaaclab.envs import (
    DirectMARLEnv,
    DirectMARLEnvCfg,
    DirectRLEnvCfg,
    ManagerBasedRLEnvCfg,
    multi_agent_to_single_agent,
)
from isaaclab.utils.dict import print_dict
from isaaclab.utils.io import dump_yaml
from isaaclab_rl.rsl_rl import RslRlOnPolicyRunnerCfg, RslRlVecEnvWrapper
from isaaclab_tasks.utils import get_checkpoint_path
from isaaclab_tasks.utils.hydra import hydra_task_config

# Import extensions to set up environment tasks
import engineai_lab.tasks  # noqa: F401
from rsl_rl.runners.on_policy_runner import OnPolicyRunner

torch.backends.cuda.matmul.allow_tf32 = True
torch.backends.cudnn.allow_tf32 = True
torch.backends.cudnn.deterministic = False
torch.backends.cudnn.benchmark = False


def _restore_training_progress(env, runner, agent_cfg, checkpoint_infos):
    """Restore environment-side curriculum state after a full RSL-RL resume."""
    base_env = env.unwrapped
    engineai_info = checkpoint_infos.get("engineai_lab", {}) if isinstance(checkpoint_infos, dict) else {}
    saved_task = engineai_info.get("task")
    if saved_task is not None and saved_task != args_cli.task:
        raise ValueError(
            f"Checkpoint belongs to task {saved_task!r}, not {args_cli.task!r}. "
            "Use --load_mode finetune for cross-task weight transfer."
        )

    if "common_step_counter" in engineai_info:
        common_step_counter = int(engineai_info["common_step_counter"])
        state_source = "checkpoint metadata"
    else:
        common_step_counter = (runner.current_learning_iteration + 1) * agent_cfg.num_steps_per_env
        state_source = "legacy checkpoint iteration estimate"

    base_env.common_step_counter = common_step_counter
    base_env.curriculum_manager.compute(env_ids=None)
    base_env.command_manager.reset(env_ids=None)
    print(f"[INFO] Restored common_step_counter={common_step_counter} from {state_source}.")
    command_term = base_env.command_manager.get_term("base_velocity")
    if hasattr(command_term.cfg.ranges, "lin_vel_x"):
        print(f"[INFO] Restored forward command range: {command_term.cfg.ranges.lin_vel_x}")


def _attach_training_state_to_checkpoints(env, runner):
    """Make RSL-RL's periodic saves include the environment curriculum counter."""
    base_env = env.unwrapped
    original_save = runner.save

    def save_with_training_state(path, infos=None):
        checkpoint_infos = dict(infos) if isinstance(infos, dict) else {}
        checkpoint_infos["engineai_lab"] = {
            "task": args_cli.task,
            "common_step_counter": int(base_env.common_step_counter),
        }
        original_save(path, infos=checkpoint_infos)

    runner.save = save_with_training_state


@hydra_task_config(args_cli.task, "rsl_rl_cfg_entry_point")
def main(env_cfg: ManagerBasedRLEnvCfg | DirectRLEnvCfg | DirectMARLEnvCfg, agent_cfg: RslRlOnPolicyRunnerCfg):
    """Train with RSL-RL agent."""
    # override configurations with non-hydra CLI arguments
    agent_cfg = cli_args.update_rsl_rl_cfg(agent_cfg, args_cli)
    env_cfg.scene.num_envs = args_cli.num_envs if args_cli.num_envs is not None else env_cfg.scene.num_envs
    agent_cfg.max_iterations = (
        args_cli.max_iterations if args_cli.max_iterations is not None else agent_cfg.max_iterations
    )

    # set the environment seed
    # note: certain randomizations occur in the environment initialization so we set the seed here
    env_cfg.seed = agent_cfg.seed
    env_cfg.sim.device = args_cli.device if args_cli.device is not None else env_cfg.sim.device
    if args_cli.distributed and args_cli.device is not None and "cpu" in args_cli.device:
        raise ValueError(
            "Distributed training is not supported when using CPU device. "
            "Please use GPU device (e.g. --device cuda) for distributed training."
        )

    if args_cli.distributed:
        env_cfg.sim.device = f"cuda:{app_launcher.local_rank}"
        agent_cfg.device = f"cuda:{app_launcher.local_rank}"

        seed = agent_cfg.seed + app_launcher.local_rank
        env_cfg.seed = seed
        agent_cfg.seed = seed

    # specify directory for logging experiments
    log_root_path = os.path.join("logs", "rsl_rl", agent_cfg.experiment_name)
    log_root_path = os.path.abspath(log_root_path)
    print(f"[INFO] Logging experiment in directory: {log_root_path}")
    # specify directory for logging runs: {time-stamp}_{run_name}
    log_dir = datetime.now().strftime("%Y-%m-%d_%H-%M-%S")
    if agent_cfg.run_name:
        log_dir += f"_{agent_cfg.run_name}"
    log_dir = os.path.join(log_root_path, log_dir)

    # set the log directory for the environment (works for all environment types)
    env_cfg.log_dir = log_dir

    # create isaac environment
    env = gym.make(args_cli.task, cfg=env_cfg, render_mode=None)
    # wrap for video recording

    # convert to single-agent instance if required by the RL algorithm
    if isinstance(env.unwrapped, DirectMARLEnv):
        env = multi_agent_to_single_agent(env)

    # wrap around environment for rsl-rl
    env = RslRlVecEnvWrapper(env)

    # create runner from rsl-rl
    runner = OnPolicyRunner(
        env, agent_cfg.to_dict(), log_dir=log_dir, device=agent_cfg.device
    )
    # write git state to logs
    runner.add_git_repo_to_log(__file__)
    # save resume path before creating a new log_dir
    if agent_cfg.resume:
        # get path to previous checkpoint
        resume_path = get_checkpoint_path(log_root_path, agent_cfg.load_run, agent_cfg.load_checkpoint)
        print(f"[INFO]: Loading model checkpoint from: {resume_path}")
        if args_cli.load_mode == "finetune":
            runner.load(
                resume_path,
                load_cfg={
                    "actor": True,
                    "critic": True,
                    "optimizer": False,
                    "iteration": False,
                    "rnd": False,
                },
            )
            print("[INFO] Loaded actor/critic weights with a fresh optimizer and curriculum.")
        else:
            checkpoint_infos = runner.load(resume_path)
            _restore_training_progress(env, runner, agent_cfg, checkpoint_infos)

        print(f"[INFO] Optimizer learning rate after load: {runner.alg.learning_rate:.6g}")

    _attach_training_state_to_checkpoints(env, runner)

    # dump the configuration into log-directory
    dump_yaml(os.path.join(log_dir, "params", "env.yaml"), env_cfg)
    dump_yaml(os.path.join(log_dir, "params", "agent.yaml"), agent_cfg)
    # dump_pickle(os.path.join(log_dir, "params", "env.pkl"), env_cfg)
    # dump_pickle(os.path.join(log_dir, "params", "agent.pkl"), agent_cfg)

    # run training
    runner.learn(num_learning_iterations=agent_cfg.max_iterations, init_at_random_ep_len=True)

    # close the simulator
    env.close()


if __name__ == "__main__":
    # run the main function
    main()
    # close sim app
    simulation_app.close()