TRACE Open-Sourced: Migrating G1 Controller to Full-Size Humanoid H2 in 8 Hours on One GPU
zhengyiluo · x · 2026-10-18
The HKUST(GZ) team has open-sourced the H2 SONIC whole-body controller and released TRACE, a new work enabling neural whole-body control transfer across humanoid embodiments.
- Key result: adapts the SONIC controller pretrained on Unitree G1 to the full-size humanoid H2, requiring only a single RTX PRO 6000 GPU for about 8 hours, with no need to train a new embodiment from scratch
- Method: combines joint and dynamics alignment, motion data curation, distillation, and LoRA-based reinforcement learning for efficient cross-embodiment transfer
- The adapted controller supports dynamic motion tracking and VR whole-body teleoperation; data collected via teleoperation can be directly used to train VLA (vision-language-action) autonomous tasks
- Paper: arXiv 2610.17999; project page and H2 deployment code are both publicly available
More from Embodied
- Fishing in VR while watching your real TV: passthrough makes mixed reality living-room setups work — Scobleizer · 2026-10-18
- Snabbit could make huge money selling household robotics training data — AashaySachdeva · 2026-10-18
- Apple's immersive horror short Remnants wows Vision Pro users with 3D presence — Scobleizer · 2026-10-18
- Humanoid Robot Funding Tops $3B as Brian Potter Explains the VLA Models Driving Robots — scottleibrand · 2026-10-18
- MIT's SANDO Drone Planner Avoids Every Moving Obstacle Across 12 Real Flights — Olivier__OG · 2026-10-18
- SimInfinity: open-source robotic simulation benchmark with 10K real-world tasks — joemeno · 2026-10-18