Open Source Robotics

A curated collection of high-quality open source robotics projects, tools, and software to propel the robotics community forward.

LeRobot now drives the Unitree G1 humanoid through a VLA and a whole-body controller

  • Source: Hugging Face LeRobot team
  • Author: Open Source Robotics
  • Published: 2026-09-25
  • Category: business

The LeRobot team has published a full humanoid workflow for the Unitree G1. A π0.5 policy predicts motion tokens and NVIDIA's SONIC controller turns them into balanced whole-body motion, with the teleoperation rig, printable hardware, dataset and weights all released. The SONIC path is on LeRobot's main branch, not yet in a release.

LeRobot can now run a learned policy on a full-size humanoid. In a post on 25 September, the LeRobot team at Hugging Face shows the whole chain on a Unitree G1: whole-body teleoperation, a dataset, a fine-tuned vision-language-action model, and a controller that keeps a 29-joint biped upright while the model decides what to do.

The design answers a problem arms never had. A VLA is far too slow to balance a walking robot on its own, so LeRobot puts a fast controller between the two. The policy is OpenPI's π0.5, and it predicts a compact 64-dimensional motion token rather than joint angles. The decoder of NVIDIA's SONIC, part of GR00T-WholeBodyControl, turns each token plus recent body state into joint targets at 50 Hz on the robot. The recipe follows the OpenHLM paper, with one change: LeRobot trains the VLA on the latent token itself. The SONIC decoder was merged into LeRobot on 17 August as a pure-Python ONNX re-implementation.

The demonstration task is carrying a can to a table. Its dataset holds about 100 episodes, roughly 71 minutes of VR teleoperation with three cameras. π0.5 was fine-tuned on it for 12,000 steps on four H100s, and the resulting policy is public. The robot itself is modified: OpenArm grippers on printed wrist adapters, a head camera, and a printed backpack for the CAN-FD adapter and cabling. The printable parts are released too.

The post also covers a second experiment and the wider stack. The team trained a LoRA adapter on SONIC in mjlab with PPO, so the G1 dodges thrown balls using its head-camera depth. That follows ViBE. In simulation it dodged 79.1% of throws, against 0% with the depth input blanked. The team says plainly that this is a simulation figure, and that the hardware clip is a demonstration, not an evaluation. The G1 can also be driven through GR00T whole-body control or Amazon FAR's Holosoma. It has a simulation in MuJoCo and arm IK, and the team's Homunculus glove and exoskeleton record teleoperation directly from joint motion. Those devices were first built for HOPEJr.

One thing to be clear about. G1 support has been in LeRobot releases for months, but the SONIC decoder merged after v0.6.1, the latest tagged release. Running the workflow described here means installing LeRobot from source, and the G1 documentation is written against the main branch. The team says the next humanoid it will support is Asimov, Menlo Research's open-source robot.

Why it matters

Humanoids are where the robot-learning field is putting its effort, and until now the open tooling mostly stopped at arms. The hard part on a biped is not the model but keeping it upright while the model thinks. Splitting the work between a slow policy and a fast whole-body controller, and publishing every piece of the chain, gives anyone with a G1 a reproducible starting point. It also means a task-level policy can be trained from about a hundred demonstrations, without relearning balance from scratch.

Key takeaways

  • LeRobot runs a π0.5 policy on the Unitree G1 by having it predict SONIC's 64-D motion tokens, which a fast on-robot decoder turns into 29-DoF joint targets at 50 Hz
  • The SONIC decoder is a pure-Python ONNX port merged into LeRobot on 17 August; it is on the main branch but not in v0.6.1, the latest release
  • The example can-carrying policy was trained on about 100 whole-body VR teleoperation episodes (roughly 71 minutes), and the dataset and weights are public
  • Printable G1 modifications are released: OpenArm gripper wrist adapters, a head-camera holder and a backpack enclosure for the CAN-FD adapter
  • A depth-based ball-dodging adapter trained in mjlab dodged 79.1% of throws in simulation; the authors say it has no measured hardware success rate
  • The G1 can also be driven through GR00T whole-body control or Holosoma, with a MuJoCo simulation, arm IK and full proprioceptive state
  • Asimov, Menlo Research's open-source humanoid, is named as the next robot to be supported

Open-source angle

Everything the post describes is released: the SONIC decoder and G1 integration in LeRobot under Apache-2.0, the teleoperation dataset, the fine-tuned policy weights, the printable hardware modifications and the Homunculus glove and exoskeleton designs. What is not open is the robot. The Unitree G1 is a commercial humanoid, so the workflow is open software on closed hardware, and the post names Asimov as the open humanoid it will support next.

Related technologies

  • LeRobot
  • Unitree G1
  • OpenPI
  • SONIC
  • GR00T-WholeBodyControl
  • OpenHLM
  • Holosoma
  • mjlab
  • OpenArm
  • MuJoCo

Related projects

  • LeRobot
  • OpenPI
  • OpenArm
  • mjlab
  • NVIDIA Isaac GR00T
  • Holosoma
  • HOPEJr
  • Asimov v1

Links

  • Source: https://huggingface.co/blog/nepyope/bringing-humanoids-to-lerobot

This website requires JavaScript to function properly. Please enable JavaScript to explore our collection of open source robotics projects.