Website | Technical Paper | Videos

This repository contains example RL environments for the NVIDIA Isaac Gym high performance environments described in our NeurIPS 2021 Datasets and Benchmarks paper

Download the Isaac Gym Preview 4 release from the website, then follow the installation instructions in the documentation. We highly recommend using a conda environment to simplify set up.

Ensure that Isaac Gym works on your system by running one of the examples from the python/examples directory, like joint_monkey.py. Follow troubleshooting steps described in the Isaac Gym Preview 4 install instructions if you have any trouble running the samples.

Once Isaac Gym is installed and samples work within your current python environment, install this repo:

pip install -e .

We offer an easy-to-use API for creating preset vectorized environments. For more info on what a vectorized environment is and its usage, please refer to the Gym library documentation.

import isaacgym
import isaacgymenvs
import torch

num_envs = 2000

envs = isaacgymenvs.make(
	seed=0, 
	task="Ant", 
	num_envs=num_envs, 
	sim_device="cuda:0",
	rl_device="cuda:0",
)
print("Observation space is", envs.observation_space)
print("Action space is", envs.action_space)
obs = envs.reset()
for _ in range(20):
	random_actions = 2.0 * torch.rand((num_envs,) + envs.action_space.shape, device = 'cuda:0') - 1.0
	envs.step(random_actions)

To train your first policy, run this line:

python train.py task=Cartpole

Cartpole should train to the point that the pole stays upright within a few seconds of starting.

Here's another example - Ant locomotion:

python train.py task=Ant

Note that by default we show a preview window, which will usually slow down training. You can use the v key while running to disable viewer updates and allow training to proceed faster. Hit the v key again to resume viewing after a few seconds of training, once the ants have learned to run a bit better.

Use the esc key or close the viewer window to stop training early.

Alternatively, you can train headlessly, as follows:

python train.py task=Ant headless=True

Ant may take a minute or two to train a policy you can run. When running headlessly, you can stop it early using Control-C in the command line window.

Checkpoints are saved in the folder runs/EXPERIMENT_NAME/nn where EXPERIMENT_NAME defaults to the task name, but can also be overridden via the experiment argument.

To load a trained checkpoint and continue training, use the checkpoint argument:

python train.py task=Ant checkpoint=runs/Ant/nn/Ant.pth

To load a trained checkpoint and only perform inference (no training), pass test=True as an argument, along with the checkpoint name. To avoid rendering overhead, you may also want to run with fewer environments using num_envs=64:

python train.py task=Ant checkpoint=runs/Ant/nn/Ant.pth test=True num_envs=64

Note that If there are special characters such as [ or = in the checkpoint names, you will need to escape them and put quotes around the string. For example, checkpoint="./runs/Ant/nn/last_Antep\=501rew\[5981.31\].pth"

We use Hydra to manage the config. Note that this has some differences from previous incarnations in older versions of Isaac Gym.

Key arguments to the train.py script are:

• task=TASK - selects which task to use. Any of AllegroHand, AllegroHandDextremeADR, AllegroHandDextremeManualDR, AllegroKukaLSTM, AllegroKukaTwoArmsLSTM, Ant, Anymal, AnymalTerrain, BallBalance, Cartpole, FrankaCabinet, Humanoid, Ingenuity Quadcopter, ShadowHand, ShadowHandOpenAI_FF, ShadowHandOpenAI_LSTM, and Trifinger (these correspond to the config for each environment in the folder isaacgymenvs/config/task)
• train=TRAIN - selects which training config to use. Will automatically default to the correct config for the environment (ie. <TASK>PPO).
• num_envs=NUM_ENVS - selects the number of environments to use (overriding the default number of environments set in the task config).
• seed=SEED - sets a seed value for randomizations, and overrides the default seed set up in the task config
• sim_device=SIM_DEVICE_TYPE - Device used for physics simulation. Set to cuda:0 (default) to use GPU and to cpu for CPU. Follows PyTorch-like device syntax.
• rl_device=RL_DEVICE - Which device / ID to use for the RL algorithm. Defaults to cuda:0, and also follows PyTorch-like device syntax.
• graphics_device_id=GRAPHICS_DEVICE_ID - Which Vulkan graphics device ID to use for rendering. Defaults to 0. Note - this may be different from CUDA device ID, and does not follow PyTorch-like device syntax.
• pipeline=PIPELINE - Which API pipeline to use. Defaults to gpu, can also set to cpu. When using the gpu pipeline, all data stays on the GPU and everything runs as fast as possible. When using the cpu pipeline, simulation can run on either CPU or GPU, depending on the sim_device setting, but a copy of the data is always made on the CPU at every step.
• test=TEST- If set to True, only runs inference on the policy and does not do any training.
• checkpoint=CHECKPOINT_PATH - Set to path to the checkpoint to load for training or testing.
• headless=HEADLESS - Whether to run in headless mode.
• experiment=EXPERIMENT - Sets the name of the experiment.
• max_iterations=MAX_ITERATIONS - Sets how many iterations to run for. Reasonable defaults are provided for the provided environments.

Hydra also allows setting variables inside config files directly as command line arguments. As an example, to set the discount rate for a rl_games training run, you can use train.params.config.gamma=0.999. Similarly, variables in task configs can also be set. For example, task.env.enableDebugVis=True.

Default values for each of these are found in the isaacgymenvs/config/config.yaml file.

The way that the task and train portions of the config works are through the use of config groups. You can learn more about how these work here The actual configs for task are in isaacgymenvs/config/task/<TASK>.yaml and for train in isaacgymenvs/config/train/<TASK>PPO.yaml.

In some places in the config you will find other variables referenced (for example, num_actors: ${....task.env.numEnvs}). Each . represents going one level up in the config hierarchy. This is documented fully here.

Source code for tasks can be found in isaacgymenvs/tasks.

Each task subclasses the VecEnv base class in isaacgymenvs/base/vec_task.py.

Refer to docs/framework.md for how to create your own tasks.

Full details on each of the tasks available can be found in the RL examples documentation.

IsaacGymEnvs includes a framework for Domain Randomization to improve Sim-to-Real transfer of trained RL policies. You can read more about it here.

If deterministic training of RL policies is important for your work, you may wish to review our Reproducibility and Determinism Documentation.

You can run multi-GPU training using torchrun (i.e., torch.distributed) using this repository.

Here is an example command for how to run in this way - torchrun --standalone --nnodes=1 --nproc_per_node=2 train.py multi_gpu=True task=Ant <OTHER_ARGS>

Where the --nproc_per_node= flag specifies how many processes to run and note the multi_gpu=True flag must be set on the train script in order for multi-GPU training to run.

You can run population based training to help find good hyperparameters or to train on very difficult environments which would otherwise be hard to learn anything on without it. See the readme for details.

You can run WandB with Isaac Gym Envs by setting wandb_activate=True flag from the command line. You can set the group, name, entity, and project for the run by setting the wandb_group, wandb_name, wandb_entity and wandb_project set. Make sure you have WandB installed with pip install wandb before activating.

We implement the standard env.render(mode='rgb_rray') gym API to provide an image of the simulator viewer. Additionally, we can leverage gym.wrappers.RecordVideo to help record videos that shows agent's gameplay. Consider running the following file which should produce a video in the videos folder.

import gym
import isaacgym
import isaacgymenvs
import torch

num_envs = 64

envs = isaacgymenvs.make(
	seed=0, 
	task="Ant", 
	num_envs=num_envs, 
	sim_device="cuda:0",
	rl_device="cuda:0",
	graphics_device_id=0,
	headless=False,
	multi_gpu=False,
	virtual_screen_capture=True,
	force_render=False,
)
envs.is_vector_env = True
envs = gym.wrappers.RecordVideo(
	envs,
	"./videos",
	step_trigger=lambda step: step % 10000 == 0, # record the videos every 10000 steps
	video_length=100  # for each video record up to 100 steps
)
envs.reset()
print("the image of Isaac Gym viewer is an array of shape", envs.render(mode="rgb_array").shape)
for _ in range(100):
	actions = 2.0 * torch.rand((num_envs,) + envs.action_space.shape, device = 'cuda:0') - 1.0
	envs.step(actions)

You can automatically capture the videos of the agents gameplay by toggling the capture_video=True flag and tune the capture frequency capture_video_freq=1500 and video length via capture_video_len=100. You can set force_render=False to disable rendering when the videos are not captured.

python train.py capture_video=True capture_video_freq=1500 capture_video_len=100 force_render=False

You can also automatically upload the videos to Weights and Biases:

python train.py task=Ant wandb_activate=True wandb_entity=nvidia wandb_project=rl_games capture_video=True force_render=False

We use pre-commit to helps us automate short tasks that improve code quality. Before making a commit to the repository, please ensure pre-commit run --all-files runs without error.

Please review the Isaac Gym installation instructions first if you run into any issues.

You can either submit issues through GitHub or through the Isaac Gym forum here.

Please cite this work as:

@misc{makoviychuk2021isaac,
      title={Isaac Gym: High Performance GPU-Based Physics Simulation For Robot Learning}, 
      author={Viktor Makoviychuk and Lukasz Wawrzyniak and Yunrong Guo and Michelle Lu and Kier Storey and Miles Macklin and David Hoeller and Nikita Rudin and Arthur Allshire and Ankur Handa and Gavriel State},
      year={2021},
      journal={arXiv preprint arXiv:2108.10470}
}

Note if you use the DexPBT: Scaling up Dexterous Manipulation for Hand-Arm Systems with Population Based Training work or the code related to Population Based Training, please cite the following paper:

@inproceedings{
	petrenko2023dexpbt,
	author = {Aleksei Petrenko, Arthur Allshire, Gavriel State, Ankur Handa, Viktor Makoviychuk},
	title = {DexPBT: Scaling up Dexterous Manipulation for Hand-Arm Systems with Population Based Training},
	booktitle = {RSS},
	year = {2023}
}

Note if you use the DeXtreme: Transfer of Agile In-hand Manipulation from Simulation to Reality work or the code related to Automatic Domain Randomisation, please cite the following paper:

@inproceedings{
	handa2023dextreme,
	author = {Ankur Handa, Arthur Allshire, Viktor Makoviychuk, Aleksei Petrenko, Ritvik Singh, Jingzhou Liu, Denys Makoviichuk, Karl Van Wyk, Alexander Zhurkevich, Balakumar Sundaralingam, Yashraj Narang, Jean-Francois Lafleche, Dieter Fox, Gavriel State},
	title = {DeXtreme: Transfer of Agile In-hand Manipulation from Simulation to Reality},
	booktitle = {ICRA},
	year = {2023}
} 

Note if you use the ANYmal rough terrain environment in your work, please ensure you cite the following work:

@misc{rudin2021learning,
      title={Learning to Walk in Minutes Using Massively Parallel Deep Reinforcement Learning}, 
      author={Nikita Rudin and David Hoeller and Philipp Reist and Marco Hutter},
      year={2021},
      journal = {arXiv preprint arXiv:2109.11978}
}

Note if you use the Trifinger environment in your work, please ensure you cite the following work:

@misc{isaacgym-trifinger,
  title     = {{Transferring Dexterous Manipulation from GPU Simulation to a Remote Real-World TriFinger}},
  author    = {Allshire, Arthur and Mittal, Mayank and Lodaya, Varun and Makoviychuk, Viktor and Makoviichuk, Denys and Widmaier, Felix and Wuthrich, Manuel and Bauer, Stefan and Handa, Ankur and Garg, Animesh},
  year      = {2021},
  journal = {arXiv preprint arXiv:2108.09779}
}

Note if you use the AMP: Adversarial Motion Priors environment in your work, please ensure you cite the following work:

@article{
	2021-TOG-AMP,
	author = {Peng, Xue Bin and Ma, Ze and Abbeel, Pieter and Levine, Sergey and Kanazawa, Angjoo},
	title = {AMP: Adversarial Motion Priors for Stylized Physics-Based Character Control},
	journal = {ACM Trans. Graph.},
	issue_date = {August 2021},
	volume = {40},
	number = {4},
	month = jul,
	year = {2021},
	articleno = {1},
	numpages = {15},
	url = {http://doi.acm.org/10.1145/3450626.3459670},
	doi = {10.1145/3450626.3459670},
	publisher = {ACM},
	address = {New York, NY, USA},
	keywords = {motion control, physics-based character animation, reinforcement learning},
} 

Note if you use the Factory simulation methods (e.g., SDF collisions, contact reduction) or Factory learning tools (e.g., assets, environments, or controllers) in your work, please cite the following paper:

@inproceedings{
	narang2022factory,
	author = {Yashraj Narang and Kier Storey and Iretiayo Akinola and Miles Macklin and Philipp Reist and Lukasz Wawrzyniak and Yunrong Guo and Adam Moravanszky and Gavriel State and Michelle Lu and Ankur Handa and Dieter Fox},
	title = {Factory: Fast contact for robotic assembly},
	booktitle = {Robotics: Science and Systems},
	year = {2022}
} 

Note if you use the IndustReal training environments or algorithms in your work, please cite the following paper:

@inproceedings{
	tang2023industreal,
	author = {Bingjie Tang and Michael A Lin and Iretiayo Akinola and Ankur Handa and Gaurav S Sukhatme and Fabio Ramos and Dieter Fox and Yashraj Narang},
	title = {IndustReal: Transferring contact-rich assembly tasks from simulation to reality},
	booktitle = {Robotics: Science and Systems},
	year = {2023}
}