initial commit

scripts/rsl_rl/cli_args.py

@@ -0,0 +1,91 @@
+# Copyright (c) 2022-2025, The Isaac Lab Project Developers.
+# All rights reserved.
+#
+# SPDX-License-Identifier: BSD-3-Clause
+
+from __future__ import annotations
+
+import argparse
+import random
+from typing import TYPE_CHECKING
+
+if TYPE_CHECKING:
+    from isaaclab_rl.rsl_rl import RslRlOnPolicyRunnerCfg
+
+
+def add_rsl_rl_args(parser: argparse.ArgumentParser):
+    """Add RSL-RL arguments to the parser.
+
+    Args:
+        parser: The parser to add the arguments to.
+    """
+    # create a new argument group
+    arg_group = parser.add_argument_group("rsl_rl", description="Arguments for RSL-RL agent.")
+    # -- experiment arguments
+    arg_group.add_argument(
+        "--experiment_name", type=str, default=None, help="Name of the experiment folder where logs will be stored."
+    )
+    arg_group.add_argument("--run_name", type=str, default=None, help="Run name suffix to the log directory.")
+    # -- load arguments
+    arg_group.add_argument("--resume", action="store_true", default=False, help="Whether to resume from a checkpoint.")
+    arg_group.add_argument("--load_run", type=str, default=None, help="Name of the run folder to resume from.")
+    arg_group.add_argument("--checkpoint", type=str, default=None, help="Checkpoint file to resume from.")
+    # -- logger arguments
+    arg_group.add_argument(
+        "--logger", type=str, default=None, choices={"wandb", "tensorboard", "neptune"}, help="Logger module to use."
+    )
+    arg_group.add_argument(
+        "--log_project_name", type=str, default=None, help="Name of the logging project when using wandb or neptune."
+    )
+
+
+def parse_rsl_rl_cfg(task_name: str, args_cli: argparse.Namespace) -> RslRlOnPolicyRunnerCfg:
+    """Parse configuration for RSL-RL agent based on inputs.
+
+    Args:
+        task_name: The name of the environment.
+        args_cli: The command line arguments.
+
+    Returns:
+        The parsed configuration for RSL-RL agent based on inputs.
+    """
+    from isaaclab_tasks.utils.parse_cfg import load_cfg_from_registry
+
+    # load the default configuration
+    rslrl_cfg: RslRlOnPolicyRunnerCfg = load_cfg_from_registry(task_name, "rsl_rl_cfg_entry_point")
+    rslrl_cfg = update_rsl_rl_cfg(rslrl_cfg, args_cli)
+    return rslrl_cfg
+
+
+def update_rsl_rl_cfg(agent_cfg: RslRlOnPolicyRunnerCfg, args_cli: argparse.Namespace):
+    """Update configuration for RSL-RL agent based on inputs.
+
+    Args:
+        agent_cfg: The configuration for RSL-RL agent.
+        args_cli: The command line arguments.
+
+    Returns:
+        The updated configuration for RSL-RL agent based on inputs.
+    """
+    # override the default configuration with CLI arguments
+    if hasattr(args_cli, "seed") and args_cli.seed is not None:
+        # randomly sample a seed if seed = -1
+        if args_cli.seed == -1:
+            args_cli.seed = random.randint(0, 10000)
+        agent_cfg.seed = args_cli.seed
+    if args_cli.resume is not None:
+        agent_cfg.resume = args_cli.resume
+    if args_cli.load_run is not None:
+        agent_cfg.load_run = args_cli.load_run
+    if args_cli.checkpoint is not None:
+        agent_cfg.load_checkpoint = args_cli.checkpoint
+    if args_cli.run_name is not None:
+        agent_cfg.run_name = args_cli.run_name
+    if args_cli.logger is not None:
+        agent_cfg.logger = args_cli.logger
+    # set the project name for wandb and neptune
+    if agent_cfg.logger in {"wandb", "neptune"} and args_cli.log_project_name:
+        agent_cfg.wandb_project = args_cli.log_project_name
+        agent_cfg.neptune_project = args_cli.log_project_name
+
+    return agent_cfg

scripts/rsl_rl/play.py

@@ -0,0 +1,170 @@
+# Copyright (c) 2022-2025, The Isaac Lab Project Developers.
+# All rights reserved.
+#
+# SPDX-License-Identifier: BSD-3-Clause
+
+"""Script to play a checkpoint if an RL agent from RSL-RL."""
+
+"""Launch Isaac Sim Simulator first."""
+
+import argparse
+
+from isaaclab.app import AppLauncher
+
+# local imports
+import cli_args  # isort: skip
+
+# add argparse arguments
+parser = argparse.ArgumentParser(description="Train an RL agent with RSL-RL.")
+parser.add_argument("--video", action="store_true", default=False, help="Record videos during training.")
+parser.add_argument("--video_length", type=int, default=200, help="Length of the recorded video (in steps).")
+parser.add_argument(
+    "--disable_fabric", action="store_true", default=False, help="Disable fabric and use USD I/O operations."
+)
+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(
+    "--use_pretrained_checkpoint",
+    action="store_true",
+    help="Use the pre-trained checkpoint from Nucleus.",
+)
+parser.add_argument("--real-time", action="store_true", default=False, help="Run in real-time, if possible.")
+# append RSL-RL cli arguments
+cli_args.add_rsl_rl_args(parser)
+# append AppLauncher cli args
+AppLauncher.add_app_launcher_args(parser)
+args_cli = parser.parse_args()
+# always enable cameras to record video
+if args_cli.video:
+    args_cli.enable_cameras = True
+
+# launch omniverse app
+app_launcher = AppLauncher(args_cli)
+simulation_app = app_launcher.app
+
+"""Rest everything follows."""
+
+import gymnasium as gym
+import os
+import time
+import torch
+
+from rsl_rl.runners import OnPolicyRunner
+
+from isaaclab.envs import DirectMARLEnv, multi_agent_to_single_agent
+from isaaclab.utils.assets import retrieve_file_path
+from isaaclab.utils.dict import print_dict
+from isaaclab.utils.pretrained_checkpoint import get_published_pretrained_checkpoint
+
+from isaaclab_rl.rsl_rl import RslRlOnPolicyRunnerCfg, RslRlVecEnvWrapper, export_policy_as_jit, export_policy_as_onnx
+
+import isaaclab_tasks  # noqa: F401
+from isaaclab_tasks.utils import get_checkpoint_path, parse_env_cfg
+
+import FLEXR_v0.tasks  # noqa: F401
+
+
+def main():
+    """Play with RSL-RL agent."""
+    # parse configuration
+    env_cfg = parse_env_cfg(
+        args_cli.task, device=args_cli.device, num_envs=args_cli.num_envs, use_fabric=not args_cli.disable_fabric
+    )
+    agent_cfg: RslRlOnPolicyRunnerCfg = cli_args.parse_rsl_rl_cfg(args_cli.task, args_cli)
+
+    # 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] Loading experiment from directory: {log_root_path}")
+    if args_cli.use_pretrained_checkpoint:
+        resume_path = get_published_pretrained_checkpoint("rsl_rl", args_cli.task)
+        if not resume_path:
+            print("[INFO] Unfortunately a pre-trained checkpoint is currently unavailable for this task.")
+            return
+    elif args_cli.checkpoint:
+        resume_path = retrieve_file_path(args_cli.checkpoint)
+    else:
+        resume_path = get_checkpoint_path(log_root_path, agent_cfg.load_run, agent_cfg.load_checkpoint)
+
+    log_dir = os.path.dirname(resume_path)
+
+    # create isaac environment
+    env = gym.make(args_cli.task, cfg=env_cfg, render_mode="rgb_array" if args_cli.video else None)
+
+    # convert to single-agent instance if required by the RL algorithm
+    if isinstance(env.unwrapped, DirectMARLEnv):
+        env = multi_agent_to_single_agent(env)
+
+    # wrap for video recording
+    if args_cli.video:
+        video_kwargs = {
+            "video_folder": os.path.join(log_dir, "videos", "play"),
+            "step_trigger": lambda step: step == 0,
+            "video_length": args_cli.video_length,
+            "disable_logger": True,
+        }
+        print("[INFO] Recording videos during training.")
+        print_dict(video_kwargs, nesting=4)
+        env = gym.wrappers.RecordVideo(env, **video_kwargs)
+
+    # wrap around environment for rsl-rl
+    env = RslRlVecEnvWrapper(env, clip_actions=agent_cfg.clip_actions)
+
+    print(f"[INFO]: Loading model checkpoint from: {resume_path}")
+    # load previously trained model
+    ppo_runner = OnPolicyRunner(env, agent_cfg.to_dict(), log_dir=None, device=agent_cfg.device)
+    ppo_runner.load(resume_path)
+
+    # obtain the trained policy for inference
+    policy = ppo_runner.get_inference_policy(device=env.unwrapped.device)
+
+    # extract the neural network module
+    # we do this in a try-except to maintain backwards compatibility.
+    try:
+        # version 2.3 onwards
+        policy_nn = ppo_runner.alg.policy
+    except AttributeError:
+        # version 2.2 and below
+        policy_nn = ppo_runner.alg.actor_critic
+
+    # export policy to onnx/jit
+    export_model_dir = os.path.join(os.path.dirname(resume_path), "exported")
+    export_policy_as_jit(policy_nn, ppo_runner.obs_normalizer, path=export_model_dir, filename="policy.pt")
+    export_policy_as_onnx(
+        policy_nn, normalizer=ppo_runner.obs_normalizer, path=export_model_dir, filename="policy.onnx"
+    )
+
+    dt = env.unwrapped.step_dt
+
+    # reset environment
+    obs, _ = env.get_observations()
+    timestep = 0
+    # simulate environment
+    while simulation_app.is_running():
+        start_time = time.time()
+        # run everything in inference mode
+        with torch.inference_mode():
+            # agent stepping
+            actions = policy(obs)
+            # env stepping
+            obs, _, _, _ = env.step(actions)
+        if args_cli.video:
+            timestep += 1
+            # Exit the play loop after recording one video
+            if timestep == args_cli.video_length:
+                break
+
+        # time delay for real-time evaluation
+        sleep_time = dt - (time.time() - start_time)
+        if args_cli.real_time and sleep_time > 0:
+            time.sleep(sleep_time)
+
+    # close the simulator
+    env.close()
+
+
+if __name__ == "__main__":
+    # run the main function
+    main()
+    # close sim app
+    simulation_app.close()

scripts/rsl_rl/train.py

@@ -0,0 +1,193 @@
+# Copyright (c) 2022-2025, The Isaac Lab Project Developers.
+# All rights reserved.
+#
+# SPDX-License-Identifier: BSD-3-Clause
+
+"""Script to train RL agent with RSL-RL."""
+
+"""Launch Isaac Sim Simulator first."""
+
+import argparse
+import sys
+
+from isaaclab.app import AppLauncher
+
+# local imports
+import cli_args  # isort: skip
+
+
+# add argparse arguments
+parser = argparse.ArgumentParser(description="Train an RL agent with RSL-RL.")
+parser.add_argument("--video", action="store_true", default=False, help="Record videos during training.")
+parser.add_argument("--video_length", type=int, default=200, help="Length of the recorded video (in steps).")
+parser.add_argument("--video_interval", type=int, default=2000, help="Interval between video recordings (in steps).")
+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()
+
+# always enable cameras to record video
+if args_cli.video:
+    args_cli.enable_cameras = True
+
+# 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
+
+"""Check for minimum supported RSL-RL version."""
+
+import importlib.metadata as metadata
+import platform
+
+from packaging import version
+
+# for distributed training, check minimum supported rsl-rl version
+RSL_RL_VERSION = "2.3.1"
+installed_version = metadata.version("rsl-rl-lib")
+if args_cli.distributed and version.parse(installed_version) < version.parse(RSL_RL_VERSION):
+    if platform.system() == "Windows":
+        cmd = [r".\isaaclab.bat", "-p", "-m", "pip", "install", f"rsl-rl-lib=={RSL_RL_VERSION}"]
+    else:
+        cmd = ["./isaaclab.sh", "-p", "-m", "pip", "install", f"rsl-rl-lib=={RSL_RL_VERSION}"]
+    print(
+        f"Please install the correct version of RSL-RL.\nExisting version is: '{installed_version}'"
+        f" and required version is: '{RSL_RL_VERSION}'.\nTo install the correct version, run:"
+        f"\n\n\t{' '.join(cmd)}\n"
+    )
+    exit(1)
+
+"""Rest everything follows."""
+
+import gymnasium as gym
+import os
+import torch
+from datetime import datetime
+
+from rsl_rl.runners import OnPolicyRunner
+
+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_pickle, dump_yaml
+
+from isaaclab_rl.rsl_rl import RslRlOnPolicyRunnerCfg, RslRlVecEnvWrapper
+
+import isaaclab_tasks  # noqa: F401
+from isaaclab_tasks.utils import get_checkpoint_path
+from isaaclab_tasks.utils.hydra import hydra_task_config
+
+import FLEXR_v0.tasks  # noqa: F401
+
+torch.backends.cuda.matmul.allow_tf32 = True
+torch.backends.cudnn.allow_tf32 = True
+torch.backends.cudnn.deterministic = False
+torch.backends.cudnn.benchmark = False
+
+
+@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
+
+    # multi-gpu training configuration
+    if args_cli.distributed:
+        env_cfg.sim.device = f"cuda:{app_launcher.local_rank}"
+        agent_cfg.device = f"cuda:{app_launcher.local_rank}"
+
+        # set seed to have diversity in different threads
+        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")
+    # The Ray Tune workflow extracts experiment name using the logging line below, hence, do not change it (see PR #2346, comment-2819298849)
+    print(f"Exact experiment name requested from command line: {log_dir}")
+    if agent_cfg.run_name:
+        log_dir += f"_{agent_cfg.run_name}"
+    log_dir = os.path.join(log_root_path, log_dir)
+
+    # create isaac environment
+    env = gym.make(args_cli.task, cfg=env_cfg, render_mode="rgb_array" if args_cli.video else None)
+
+    # convert to single-agent instance if required by the RL algorithm
+    if isinstance(env.unwrapped, DirectMARLEnv):
+        env = multi_agent_to_single_agent(env)
+
+    # save resume path before creating a new log_dir
+    if agent_cfg.resume or agent_cfg.algorithm.class_name == "Distillation":
+        resume_path = get_checkpoint_path(log_root_path, agent_cfg.load_run, agent_cfg.load_checkpoint)
+
+    # wrap for video recording
+    if args_cli.video:
+        video_kwargs = {
+            "video_folder": os.path.join(log_dir, "videos", "train"),
+            "step_trigger": lambda step: step % args_cli.video_interval == 0,
+            "video_length": args_cli.video_length,
+            "disable_logger": True,
+        }
+        print("[INFO] Recording videos during training.")
+        print_dict(video_kwargs, nesting=4)
+        env = gym.wrappers.RecordVideo(env, **video_kwargs)
+
+    # wrap around environment for rsl-rl
+    env = RslRlVecEnvWrapper(env, clip_actions=agent_cfg.clip_actions)
+
+    # 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__)
+    # load the checkpoint
+    if agent_cfg.resume or agent_cfg.algorithm.class_name == "Distillation":
+        print(f"[INFO]: Loading model checkpoint from: {resume_path}")
+        # load previously trained model
+        runner.load(resume_path)
+
+    # 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()