legged_gym

ETH Zürich's open-source reinforcement-learning training environment for legged robots, built on Isaac Gym.

legged_gym is an open-source reinforcement-learning codebase for legged robots from ETH Zürich's Robotic Systems Lab (RSL, Nikita Rudin and colleagues), accompanying their CoRL 2021 paper “Learning to Walk in Minutes…”. It runs thousands of simulated robots on a single GPU via Isaac Gym and trains locomotion with PPO (Proximal Policy Optimization); in the paper, an ANYmal robot learned to walk on flat ground in under 4 minutes and on rough terrain in about 20 minutes. The codebase includes the pieces needed for a policy to survive the transfer to a real robot: actuator networks, friction and mass randomization, observation noise, random pushes, and a terrain curriculum that automatically raises difficulty as the policy improves; it ships with the ANYmal B/C, A1, and Cassie robots built in. In January 2024 the authors announced a move to Isaac Lab and said the original repository would only get limited maintenance, but many projects, including Unitree's unitree_rl_gym, still use it as their backbone.

ExampleUnitree's open-source unitree_rl_gym follows the same structure as legged_gym: running python legged_gym/scripts/train.py --task=g1 trains a G1 humanoid to walk inside Isaac Gym, after which the policy is validated with sim-to-sim testing in MuJoCo before being deployed to the real robot.

Also called

Isaac Gym Environments for Legged Robots, Legged Gym

Related

Isaac Gym · rsl_rl · Terrain Curriculum · Massively Parallel Reinforcement Learning · unitree_rl_gym (Unitree RL Gym) · NVIDIA Isaac Lab

Sources

leggedrobotics/legged_gym (GitHub)
Learning to Walk in Minutes Using Massively Parallel Deep Reinforcement Learning (arXiv 2109.11978)
unitreerobotics/unitree_rl_gym (GitHub)

As of

2026-09