SonicStar
Open-source Unitree G1 Vision-Language-Action stack for teleop data collection, SonicLatent training, simulation, and real-time whole-body policy deployment(real world deployment TBD).
A training and deployment toolkit for AI-controlled motion on the Unitree G1 humanoid robot. Give the robot a plain-text instruction and it uses camera input and a trained neural network to carry out the physical task.
SonicStar is an open-source project for training and deploying AI-controlled behavior on the Unitree G1 humanoid robot. The core idea is called Vision-Language-Action (VLA): you give the robot a plain-text instruction like "pick up the cylinder and throw it into the trash bin" and it executes the physical movements needed to follow that instruction, guided by camera input and a trained neural network.
The repository is split into two main parts. The first (starVLA/) handles training and inference. You collect demonstration data from a simulation, train a model called SonicLatent on that data, then run a policy server that accepts text prompts and returns control signals. The second part (wbc/) handles the whole-body control side: running the simulation environment, capturing teleoperation data from a human operator, and connecting the trained policy to the robot's motors.
The intended workflow is: run the simulation to gather training examples of a human completing tasks, process those recordings into a dataset, train the model, then deploy it back into the simulation (and eventually to the real robot) to run the task autonomously. The README includes keyboard shortcuts used during deployment to move the simulated robot through its initialization states before the policy takes over.
The README is written entirely in Chinese and references two external codebases (NVIDIA's GR00T-WholeBodyControl and the starVLA framework) as the main systems this repository is built on top of. Real-world deployment on physical hardware is listed as a future milestone.
Where it fits
- Train an AI policy that makes a Unitree G1 humanoid robot follow text instructions by learning from simulated human demonstrations.
- Capture teleoperation data in simulation where a human operator controls the robot to build a training dataset for motion policies.
- Deploy a trained Vision-Language-Action model to run tasks autonomously in simulation before attempting transfer to real hardware.