LeRobot adds Unitree G1 remote inference via SONIC
Decision Brief
The PR adds SONIC remote inference support for the Unitree G1. Under the SonicWholeBodyController, the G1 declares a rollout wait contract: on standby it resends the last full 64-dim motion token so the onboard decoder keeps balancing. Token echo now uses the action key, so observed tokens can be replayed before the first policy result, and the echo is the decoder's neutral token rather than all-zeros (which decodes to a bent-over pose). It adds a nepyope/sonic_walk server config and a hardware-validated G1 operating procedure. The locomotion controller and raw joint control remain unsupported. For teams running remote policies on the Unitree G1 with LeRobot, deployment becomes: the policy server runs on a GPU machine, while the robot bridge, SONIC decoder, and 50Hz rt/lowcmd loop stay on the G1's Jetson, with only 64-dim token chunks sent over the network. When the action buffer runs dry, rollout repeats the last token for up to one second (controlled by action_starvation_grace_s); with no new prediction it errors out and follows the shutdown or disconnect path. The G1 locomotion controller is still unsupported, since gamepad velocity can't be preserved this way.
Sources
- LeRobot(HuggingFace 机器人)
HuggingFace LeRobot — embodied-AI / robotics learning library; new models, datasets, and policies.
- LeRobot(HuggingFace 机器人)
留言
登入后即可留言,和其他 builder 交换实测心得。
还没有留言,抢头香。