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 omni.isaac.lab_tasks.utils.wrappers 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 source/standalone/workflows/sb3 directory.

Code for train.py
  1# Copyright (c) 2022-2024, The Isaac Lab Project Developers.
  2# All rights reserved.
  3#
  4# SPDX-License-Identifier: BSD-3-Clause
  5
  6"""Script to train RL agent with Stable Baselines3.
  7
  8Since Stable-Baselines3 does not support buffers living on GPU directly,
  9we recommend using smaller number of environments. Otherwise,
 10there will be significant overhead in GPU->CPU transfer.
 11"""
 12
 13"""Launch Isaac Sim Simulator first."""
 14
 15import argparse
 16import sys
 17
 18from omni.isaac.lab.app import AppLauncher
 19
 20# add argparse arguments
 21parser = argparse.ArgumentParser(description="Train an RL agent with Stable-Baselines3.")
 22parser.add_argument("--video", action="store_true", default=False, help="Record videos during training.")
 23parser.add_argument("--video_length", type=int, default=200, help="Length of the recorded video (in steps).")
 24parser.add_argument("--video_interval", type=int, default=2000, help="Interval between video recordings (in steps).")
 25parser.add_argument("--num_envs", type=int, default=None, help="Number of environments to simulate.")
 26parser.add_argument("--task", type=str, default=None, help="Name of the task.")
 27parser.add_argument("--seed", type=int, default=None, help="Seed used for the environment")
 28parser.add_argument("--max_iterations", type=int, default=None, help="RL Policy training iterations.")
 29# append AppLauncher cli args
 30AppLauncher.add_app_launcher_args(parser)
 31# parse the arguments
 32args_cli, hydra_args = parser.parse_known_args()
 33# always enable cameras to record video
 34if args_cli.video:
 35    args_cli.enable_cameras = True
 36
 37# clear out sys.argv for Hydra
 38sys.argv = [sys.argv[0]] + hydra_args
 39
 40# launch omniverse app
 41app_launcher = AppLauncher(args_cli)
 42simulation_app = app_launcher.app
 43
 44"""Rest everything follows."""
 45
 46import gymnasium as gym
 47import numpy as np
 48import os
 49import random
 50from datetime import datetime
 51
 52from stable_baselines3 import PPO
 53from stable_baselines3.common.callbacks import CheckpointCallback
 54from stable_baselines3.common.logger import configure
 55from stable_baselines3.common.vec_env import VecNormalize
 56
 57from omni.isaac.lab.envs import (
 58    DirectMARLEnv,
 59    DirectMARLEnvCfg,
 60    DirectRLEnvCfg,
 61    ManagerBasedRLEnvCfg,
 62    multi_agent_to_single_agent,
 63)
 64from omni.isaac.lab.utils.dict import print_dict
 65from omni.isaac.lab.utils.io import dump_pickle, dump_yaml
 66
 67import omni.isaac.lab_tasks  # noqa: F401
 68from omni.isaac.lab_tasks.utils.hydra import hydra_task_config
 69from omni.isaac.lab_tasks.utils.wrappers.sb3 import Sb3VecEnvWrapper, process_sb3_cfg
 70
 71
 72@hydra_task_config(args_cli.task, "sb3_cfg_entry_point")
 73def main(env_cfg: ManagerBasedRLEnvCfg | DirectRLEnvCfg | DirectMARLEnvCfg, agent_cfg: dict):
 74    """Train with stable-baselines agent."""
 75    # randomly sample a seed if seed = -1
 76    if args_cli.seed == -1:
 77        args_cli.seed = random.randint(0, 10000)
 78
 79    # override configurations with non-hydra CLI arguments
 80    env_cfg.scene.num_envs = args_cli.num_envs if args_cli.num_envs is not None else env_cfg.scene.num_envs
 81    agent_cfg["seed"] = args_cli.seed if args_cli.seed is not None else agent_cfg["seed"]
 82    # max iterations for training
 83    if args_cli.max_iterations is not None:
 84        agent_cfg["n_timesteps"] = args_cli.max_iterations * agent_cfg["n_steps"] * env_cfg.scene.num_envs
 85
 86    # set the environment seed
 87    # note: certain randomizations occur in the environment initialization so we set the seed here
 88    env_cfg.seed = agent_cfg["seed"]
 89    env_cfg.sim.device = args_cli.device if args_cli.device is not None else env_cfg.sim.device
 90
 91    # directory for logging into
 92    log_dir = os.path.join("logs", "sb3", args_cli.task, datetime.now().strftime("%Y-%m-%d_%H-%M-%S"))
 93    # dump the configuration into log-directory
 94    dump_yaml(os.path.join(log_dir, "params", "env.yaml"), env_cfg)
 95    dump_yaml(os.path.join(log_dir, "params", "agent.yaml"), agent_cfg)
 96    dump_pickle(os.path.join(log_dir, "params", "env.pkl"), env_cfg)
 97    dump_pickle(os.path.join(log_dir, "params", "agent.pkl"), agent_cfg)
 98
 99    # post-process agent configuration
100    agent_cfg = process_sb3_cfg(agent_cfg)
101    # read configurations about the agent-training
102    policy_arch = agent_cfg.pop("policy")
103    n_timesteps = agent_cfg.pop("n_timesteps")
104
105    # create isaac environment
106    env = gym.make(args_cli.task, cfg=env_cfg, render_mode="rgb_array" if args_cli.video else None)
107    # wrap for video recording
108    if args_cli.video:
109        video_kwargs = {
110            "video_folder": os.path.join(log_dir, "videos", "train"),
111            "step_trigger": lambda step: step % args_cli.video_interval == 0,
112            "video_length": args_cli.video_length,
113            "disable_logger": True,
114        }
115        print("[INFO] Recording videos during training.")
116        print_dict(video_kwargs, nesting=4)
117        env = gym.wrappers.RecordVideo(env, **video_kwargs)
118
119    # convert to single-agent instance if required by the RL algorithm
120    if isinstance(env.unwrapped, DirectMARLEnv):
121        env = multi_agent_to_single_agent(env)
122
123    # wrap around environment for stable baselines
124    env = Sb3VecEnvWrapper(env)
125
126    if "normalize_input" in agent_cfg:
127        env = VecNormalize(
128            env,
129            training=True,
130            norm_obs="normalize_input" in agent_cfg and agent_cfg.pop("normalize_input"),
131            norm_reward="normalize_value" in agent_cfg and agent_cfg.pop("normalize_value"),
132            clip_obs="clip_obs" in agent_cfg and agent_cfg.pop("clip_obs"),
133            gamma=agent_cfg["gamma"],
134            clip_reward=np.inf,
135        )
136
137    # create agent from stable baselines
138    agent = PPO(policy_arch, env, verbose=1, **agent_cfg)
139    # configure the logger
140    new_logger = configure(log_dir, ["stdout", "tensorboard"])
141    agent.set_logger(new_logger)
142
143    # callbacks for agent
144    checkpoint_callback = CheckpointCallback(save_freq=1000, save_path=log_dir, name_prefix="model", verbose=2)
145    # train the agent
146    agent.learn(total_timesteps=n_timesteps, callback=checkpoint_callback)
147    # save the final model
148    agent.save(os.path.join(log_dir, "model"))
149
150    # close the simulator
151    env.close()
152
153
154if __name__ == "__main__":
155    # run the main function
156    main()
157    # close sim app
158    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 source/standalone/workflows/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 source/standalone/workflows/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 source/standalone/workflows/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 source/standalone/workflows/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.