LOADING

Unitree Open-Sources RL Lab for Go2, H1, and G1-29dof, from Isaac Lab to the Robot

News

Unitree has published Unitree RL Lab at https://github.com/unitreerobotics/unitreerllab . It is built on NVIDIA Isaac Lab and the repository says it supports the Go2 quadruped and the H1 and G1-29dof humanoids. A trained policy is checked in Mujoco as sim-to-sim, then deployed to the robot through unitree_sdk2. The documented example task is Unitree-G1-29dof-Velocity.

What the repository already spells out

The install notes say to set up Isaac Lab first, then clone this repo outside the Isaac Lab directory. The bundled unitreerllab.sh script installs the library, lists tasks, trains, and runs inference. Robot models can be loaded two ways. USD files come from the unitreemodel set on Hugging Face. URDF is the recommended path, and only for Isaac Sim 5.0 or newer, with files from unitreeros. Deployment means installing unitreesdk2 and building the controller in the folder for that robot. Sim-to-sim uses unitreemujoco: stand the robot up in simulation, put the feet on the ground, then run the policy.

Unitree’s open-source page also lists unitreerlgym, for Go2, H1, H1_2, and G1. The WeChat post names a third project, Unitree RL Mjlab, and does not give a repository address. The gap between a simulation score and a real grasp is covered in Fable 5.2 and GR00T 2.

Efficiency claims that are only in the post

The WeChat article adds several points the repository front page does not back with a comparison: PhysX 5.1 and TensorRT, distributed training on multiple GPUs, more than 50 percent higher training efficiency than traditional simulators, and compatibility fixes for Isaac Lab 2.3.0 and Isaac Sim 5.1.0. Those stay as the post’s claims, not as a published benchmark. It also says preset tasks cover velocity control and gait planning, plus custom tasks. The public example that matches directly is the velocity task Unitree-G1-29dof-Velocity.

The path does not remove the reality gap

The repository writes the path as training, a Mujoco check, then SDK2 on the robot. Passing the check means there is a second simulation in the middle. It does not show that the policy still holds once dust, friction, and delay are in the loop. The compute coupons, the group-chat QR code, and the “read the original” line at the end of the WeChat post are the publisher’s own promotion. They are not part of the repository.

FAQ

Q: Which robots does Unitree RL Lab support? A: The repository front page lists Go2, H1, and G1-29dof. The example training task is Unitree-G1-29dof-Velocity.

Q: How does a trained policy get onto the robot? A: Sim-to-sim in Mujoco comes first, then deployment through unitree_sdk2. URDF is the recommended model format and needs Isaac Sim 5.0 or newer.

Q: Is the “50 percent faster training” claim a published result? A: It appears in the WeChat post. The unitreerllab README does not include that comparison, so the 50 percent figure is not a published benchmark.

© 版权声明

相关文章