unitree_rl_gym (Unitree RL Gym)

Unitree's official open-source reinforcement-learning example framework, covering everything from simulated training to real-robot deployment.

unitree_rl_gym is a reinforcement-learning example repository Unitree Robotics open-sourced on GitHub, built on ETH Zurich's legged_gym (a legged-robot training framework on top of Isaac Gym) and the rsl_rl algorithm library, supporting the Go2 quadruped as well as the H1, H1_2, and G1 humanoids. It splits the workflow into four steps: train in Isaac Gym (Train), replay and check the result in simulation (Play), move the policy into MuJoCo for sim-to-sim validation (Sim2Sim), and finally deploy to the real robot (Sim2Real). The repository can export either MLP or LSTM policy networks, and provides a C++ deployment example for the G1. Unitree separately maintains a unitree_rl_lab repository built on Isaac Lab (2.3 and above), which is a different project.

ExampleA common path for newcomers: run the repo's train.py with --task=g1 to train a G1 walking policy, use play.py to replay and export the network, load it in MuJoCo to confirm it doesn't fall, then deploy it to the real robot.

Also called

Unitree RL Gym

Related

legged_gym · Isaac Gym · rsl_rl · Sim-to-Sim Transfer · Unitree G1 · MuJoCo (Multi-Joint dynamics with Contact)

Sources

unitreerobotics/unitree_rl_gym GitHub 仓库 (Chinese)
unitreerobotics/unitree_rl_lab GitHub 仓库 (Chinese)

As of

2026-09