Training with an RL Agent#

In the previous tutorials, we covered how to define an RL task environment, register it into the gym registry, and interact with it using a random agent. We now move on to the next step: training an RL agent to solve the task.

Although the envs.ManagerBasedRLEnv conforms to the gymnasium.Env interface, it is not exactly a gym environment. The input and outputs of the environment are not numpy arrays, but rather based on torch tensors with the first dimension being the number of environment instances.

Additionally, most RL libraries expect their own variation of an environment interface. For example, Stable-Baselines3 expects the environment to conform to its VecEnv API which expects a list of numpy arrays instead of a single tensor. Similarly, RSL-RL, RL-Games and SKRL expect a different interface. Since there is no one-size-fits-all solution, we do not base the envs.ManagerBasedRLEnv on any particular learning library. Instead, we implement wrappers to convert the environment into the expected interface. These are specified in the isaaclab_rl module.

In this tutorial, we will use Stable-Baselines3 to train an RL agent to solve the cartpole balancing task.

Caution

Wrapping the environment with the respective learning framework’s wrapper should happen in the end, i.e. after all other wrappers have been applied. This is because the learning framework’s wrapper modifies the interpretation of environment’s APIs which may no longer be compatible with gymnasium.Env.

The Code#

For this tutorial, we use the training script from Stable-Baselines3 workflow in the scripts/reinforcement_learning/sb3 directory.

Code for train.py
  1# Copyright (c) 2022-2025, The Isaac Lab Project Developers (https://github.com/isaac-sim/IsaacLab/blob/main/CONTRIBUTORS.md).
  2# All rights reserved.
  3#
  4# SPDX-License-Identifier: BSD-3-Clause
  5
  6
  7"""Script to train RL agent with Stable Baselines3."""
  8
  9"""Launch Isaac Sim Simulator first."""
 10
 11import argparse
 12import contextlib
 13import signal
 14import sys
 15from pathlib import Path
 16
 17from isaaclab.app import AppLauncher
 18
 19# add argparse arguments
 20parser = argparse.ArgumentParser(description="Train an RL agent with Stable-Baselines3.")
 21parser.add_argument("--video", action="store_true", default=False, help="Record videos during training.")
 22parser.add_argument("--video_length", type=int, default=200, help="Length of the recorded video (in steps).")
 23parser.add_argument("--video_interval", type=int, default=2000, help="Interval between video recordings (in steps).")
 24parser.add_argument("--num_envs", type=int, default=None, help="Number of environments to simulate.")
 25parser.add_argument("--task", type=str, default=None, help="Name of the task.")
 26parser.add_argument(
 27    "--agent", type=str, default="sb3_cfg_entry_point", help="Name of the RL agent configuration entry point."
 28)
 29parser.add_argument("--seed", type=int, default=None, help="Seed used for the environment")
 30parser.add_argument("--log_interval", type=int, default=100_000, help="Log data every n timesteps.")
 31parser.add_argument("--checkpoint", type=str, default=None, help="Continue the training from checkpoint.")
 32parser.add_argument("--max_iterations", type=int, default=None, help="RL Policy training iterations.")
 33parser.add_argument("--export_io_descriptors", action="store_true", default=False, help="Export IO descriptors.")
 34parser.add_argument(
 35    "--keep_all_info",
 36    action="store_true",
 37    default=False,
 38    help="Use a slower SB3 wrapper but keep all the extra training info.",
 39)
 40# append AppLauncher cli args
 41AppLauncher.add_app_launcher_args(parser)
 42# parse the arguments
 43args_cli, hydra_args = parser.parse_known_args()
 44# always enable cameras to record video
 45if args_cli.video:
 46    args_cli.enable_cameras = True
 47
 48# clear out sys.argv for Hydra
 49sys.argv = [sys.argv[0]] + hydra_args
 50
 51# launch omniverse app
 52app_launcher = AppLauncher(args_cli)
 53simulation_app = app_launcher.app
 54
 55
 56def cleanup_pbar(*args):
 57    """
 58    A small helper to stop training and
 59    cleanup progress bar properly on ctrl+c
 60    """
 61    import gc
 62
 63    tqdm_objects = [obj for obj in gc.get_objects() if "tqdm" in type(obj).__name__]
 64    for tqdm_object in tqdm_objects:
 65        if "tqdm_rich" in type(tqdm_object).__name__:
 66            tqdm_object.close()
 67    raise KeyboardInterrupt
 68
 69
 70# disable KeyboardInterrupt override
 71signal.signal(signal.SIGINT, cleanup_pbar)
 72
 73"""Rest everything follows."""
 74
 75import gymnasium as gym
 76import numpy as np
 77import os
 78import random
 79from datetime import datetime
 80
 81import omni
 82from stable_baselines3 import PPO
 83from stable_baselines3.common.callbacks import CheckpointCallback, LogEveryNTimesteps
 84from stable_baselines3.common.vec_env import VecNormalize
 85
 86from isaaclab.envs import (
 87    DirectMARLEnv,
 88    DirectMARLEnvCfg,
 89    DirectRLEnvCfg,
 90    ManagerBasedRLEnvCfg,
 91    multi_agent_to_single_agent,
 92)
 93from isaaclab.utils.dict import print_dict
 94from isaaclab.utils.io import dump_pickle, dump_yaml
 95
 96from isaaclab_rl.sb3 import Sb3VecEnvWrapper, process_sb3_cfg
 97
 98import isaaclab_tasks  # noqa: F401
 99from isaaclab_tasks.utils.hydra import hydra_task_config
100
101# PLACEHOLDER: Extension template (do not remove this comment)
102
103
104@hydra_task_config(args_cli.task, args_cli.agent)
105def main(env_cfg: ManagerBasedRLEnvCfg | DirectRLEnvCfg | DirectMARLEnvCfg, agent_cfg: dict):
106    """Train with stable-baselines agent."""
107    # randomly sample a seed if seed = -1
108    if args_cli.seed == -1:
109        args_cli.seed = random.randint(0, 10000)
110
111    # override configurations with non-hydra CLI arguments
112    env_cfg.scene.num_envs = args_cli.num_envs if args_cli.num_envs is not None else env_cfg.scene.num_envs
113    agent_cfg["seed"] = args_cli.seed if args_cli.seed is not None else agent_cfg["seed"]
114    # max iterations for training
115    if args_cli.max_iterations is not None:
116        agent_cfg["n_timesteps"] = args_cli.max_iterations * agent_cfg["n_steps"] * env_cfg.scene.num_envs
117
118    # set the environment seed
119    # note: certain randomizations occur in the environment initialization so we set the seed here
120    env_cfg.seed = agent_cfg["seed"]
121    env_cfg.sim.device = args_cli.device if args_cli.device is not None else env_cfg.sim.device
122
123    # directory for logging into
124    run_info = datetime.now().strftime("%Y-%m-%d_%H-%M-%S")
125    log_root_path = os.path.abspath(os.path.join("logs", "sb3", args_cli.task))
126    print(f"[INFO] Logging experiment in directory: {log_root_path}")
127    # The Ray Tune workflow extracts experiment name using the logging line below, hence, do not change it (see PR #2346, comment-2819298849)
128    print(f"Exact experiment name requested from command line: {run_info}")
129    log_dir = os.path.join(log_root_path, run_info)
130    # dump the configuration into log-directory
131    dump_yaml(os.path.join(log_dir, "params", "env.yaml"), env_cfg)
132    dump_yaml(os.path.join(log_dir, "params", "agent.yaml"), agent_cfg)
133    dump_pickle(os.path.join(log_dir, "params", "env.pkl"), env_cfg)
134    dump_pickle(os.path.join(log_dir, "params", "agent.pkl"), agent_cfg)
135
136    # save command used to run the script
137    command = " ".join(sys.orig_argv)
138    (Path(log_dir) / "command.txt").write_text(command)
139
140    # post-process agent configuration
141    agent_cfg = process_sb3_cfg(agent_cfg, env_cfg.scene.num_envs)
142    # read configurations about the agent-training
143    policy_arch = agent_cfg.pop("policy")
144    n_timesteps = agent_cfg.pop("n_timesteps")
145
146    # set the IO descriptors export flag if requested
147    if isinstance(env_cfg, ManagerBasedRLEnvCfg):
148        env_cfg.export_io_descriptors = args_cli.export_io_descriptors
149    else:
150        omni.log.warn(
151            "IO descriptors are only supported for manager based RL environments. No IO descriptors will be exported."
152        )
153
154    # set the log directory for the environment (works for all environment types)
155    env_cfg.log_dir = log_dir
156
157    # create isaac environment
158    env = gym.make(args_cli.task, cfg=env_cfg, render_mode="rgb_array" if args_cli.video else None)
159
160    # convert to single-agent instance if required by the RL algorithm
161    if isinstance(env.unwrapped, DirectMARLEnv):
162        env = multi_agent_to_single_agent(env)
163
164    # wrap for video recording
165    if args_cli.video:
166        video_kwargs = {
167            "video_folder": os.path.join(log_dir, "videos", "train"),
168            "step_trigger": lambda step: step % args_cli.video_interval == 0,
169            "video_length": args_cli.video_length,
170            "disable_logger": True,
171        }
172        print("[INFO] Recording videos during training.")
173        print_dict(video_kwargs, nesting=4)
174        env = gym.wrappers.RecordVideo(env, **video_kwargs)
175
176    # wrap around environment for stable baselines
177    env = Sb3VecEnvWrapper(env, fast_variant=not args_cli.keep_all_info)
178
179    norm_keys = {"normalize_input", "normalize_value", "clip_obs"}
180    norm_args = {}
181    for key in norm_keys:
182        if key in agent_cfg:
183            norm_args[key] = agent_cfg.pop(key)
184
185    if norm_args and norm_args.get("normalize_input"):
186        print(f"Normalizing input, {norm_args=}")
187        env = VecNormalize(
188            env,
189            training=True,
190            norm_obs=norm_args["normalize_input"],
191            norm_reward=norm_args.get("normalize_value", False),
192            clip_obs=norm_args.get("clip_obs", 100.0),
193            gamma=agent_cfg["gamma"],
194            clip_reward=np.inf,
195        )
196
197    # create agent from stable baselines
198    agent = PPO(policy_arch, env, verbose=1, tensorboard_log=log_dir, **agent_cfg)
199    if args_cli.checkpoint is not None:
200        agent = agent.load(args_cli.checkpoint, env, print_system_info=True)
201
202    # callbacks for agent
203    checkpoint_callback = CheckpointCallback(save_freq=1000, save_path=log_dir, name_prefix="model", verbose=2)
204    callbacks = [checkpoint_callback, LogEveryNTimesteps(n_steps=args_cli.log_interval)]
205
206    # train the agent
207    with contextlib.suppress(KeyboardInterrupt):
208        agent.learn(
209            total_timesteps=n_timesteps,
210            callback=callbacks,
211            progress_bar=True,
212            log_interval=None,
213        )
214    # save the final model
215    agent.save(os.path.join(log_dir, "model"))
216    print("Saving to:")
217    print(os.path.join(log_dir, "model.zip"))
218
219    if isinstance(env, VecNormalize):
220        print("Saving normalization")
221        env.save(os.path.join(log_dir, "model_vecnormalize.pkl"))
222
223    # close the simulator
224    env.close()
225
226
227if __name__ == "__main__":
228    # run the main function
229    main()
230    # close sim app
231    simulation_app.close()

The Code Explained#

Most of the code above is boilerplate code to create logging directories, saving the parsed configurations, and setting up different Stable-Baselines3 components. For this tutorial, the important part is creating the environment and wrapping it with the Stable-Baselines3 wrapper.

There are three wrappers used in the code above:

  1. gymnasium.wrappers.RecordVideo: This wrapper records a video of the environment and saves it to the specified directory. This is useful for visualizing the agent’s behavior during training.

  2. wrappers.sb3.Sb3VecEnvWrapper: This wrapper converts the environment into a Stable-Baselines3 compatible environment.

  3. stable_baselines3.common.vec_env.VecNormalize: This wrapper normalizes the environment’s observations and rewards.

Each of these wrappers wrap around the previous wrapper by following env = wrapper(env, *args, **kwargs) repeatedly. The final environment is then used to train the agent. For more information on how these wrappers work, please refer to the Wrapping environments documentation.

The Code Execution#

We train a PPO agent from Stable-Baselines3 to solve the cartpole balancing task.

Training the agent#

There are three main ways to train the agent. Each of them has their own advantages and disadvantages. It is up to you to decide which one you prefer based on your use case.

Headless execution#

If the --headless flag is set, the simulation is not rendered during training. This is useful when training on a remote server or when you do not want to see the simulation. Typically, it speeds up the training process since only physics simulation step is performed.

./isaaclab.sh -p scripts/reinforcement_learning/sb3/train.py --task Isaac-Cartpole-v0 --num_envs 64 --headless

Headless execution with off-screen render#

Since the above command does not render the simulation, it is not possible to visualize the agent’s behavior during training. To visualize the agent’s behavior, we pass the --enable_cameras which enables off-screen rendering. Additionally, we pass the flag --video which records a video of the agent’s behavior during training.

./isaaclab.sh -p scripts/reinforcement_learning/sb3/train.py --task Isaac-Cartpole-v0 --num_envs 64 --headless --video

The videos are saved to the logs/sb3/Isaac-Cartpole-v0/<run-dir>/videos/train directory. You can open these videos using any video player.

Interactive execution#

While the above two methods are useful for training the agent, they don’t allow you to interact with the simulation to see what is happening. In this case, you can ignore the --headless flag and run the training script as follows:

./isaaclab.sh -p scripts/reinforcement_learning/sb3/train.py --task Isaac-Cartpole-v0 --num_envs 64

This will open the Isaac Sim window and you can see the agent training in the environment. However, this will slow down the training process since the simulation is rendered on the screen. As a workaround, you can switch between different render modes in the "Isaac Lab" window that is docked on the bottom-right corner of the screen. To learn more about these render modes, please check the sim.SimulationContext.RenderMode class.

Viewing the logs#

On a separate terminal, you can monitor the training progress by executing the following command:

# execute from the root directory of the repository
./isaaclab.sh -p -m tensorboard.main --logdir logs/sb3/Isaac-Cartpole-v0

Playing the trained agent#

Once the training is complete, you can visualize the trained agent by executing the following command:

# execute from the root directory of the repository
./isaaclab.sh -p scripts/reinforcement_learning/sb3/play.py --task Isaac-Cartpole-v0 --num_envs 32 --use_last_checkpoint

The above command will load the latest checkpoint from the logs/sb3/Isaac-Cartpole-v0 directory. You can also specify a specific checkpoint by passing the --checkpoint flag.