Sithija Ranaraja

All projectsHumanoid robotics

Humanoid control

Unitree G1 whole-body control with GR00T and SONIC

Eight episodes from the egocentric manipulation dataset, each captioned with its task instruction.Simulation

A loco-manipulation stack for the Unitree G1. NVIDIA GR00T is the vision-language-action policy and outputs latent motion tokens. GEAR-SONIC is the whole-body controller that turns those tokens into leg, waist, arm and hand motion.

Type
Simulation
Period
2026 – present
Role
Architecture, simulation and controller development
Stack
Unitree G1, NVIDIA GR00T, GEAR-SONIC, Isaac Sim and Isaac Lab, MuJoCo, Genesis, LeRobot, unitree_sdk2_python, unitree_rl_gym, ROS 2
Result
Dataset: 82 validated episodes, 27,915 frames, 55 tasks. Push recovery: a 320 N lateral shove, and the G1 stays up.

Challenge

The policy has to understand the task and the controller has to keep the robot upright while the arms work. Neither is safe to tune on hardware first.

Approach

Everything starts in simulation: bring up the SONIC sim-to-sim loop, generate demonstrations, fine-tune the policy, evaluate, then move to hardware. Along the way I evaluated the BONES-SEED, GRAIL and OmniRetarget motion datasets and the Unitree SDK. Motion-capture retargeting drives the upper body on the simulated G1.

What I built

  • GR00T VLA with the SONIC whole-body controller on the Unitree G1
  • Pipeline in Isaac Lab and MuJoCo
  • Synthetic egocentric manipulation dataset of 82 validated episodes
  • Evaluated BONES-SEED, GRAIL and OmniRetarget as data sources
  • Motion-capture retargeting for upper-body motion

Egocentric manipulation dataset

82 validated episodes recorded from the G1's own head camera while the whole-body controller keeps it balanced. Tasks: fruit onto a plate or tray, stacking and dropping onto a block, left-hand variants, a two-handed board lift, pouring into a bowl, opening a drawer. Each episode carries its language instruction.

Manipulation under whole-body control

Arm and hand tasks run on top of the whole-body controller. It keeps the robot balanced while the policy commands the motion.

Locomotion and whole-body control

The same controller walks the robot to a goal before manipulation starts.