TRACE: Efficient Transfer Learning for Neural Whole-Body Control Across Humanoid Embodiments
TRACE: Efficient Transfer Learning for Neural Whole-Body Control Across Humanoid Embodiments
TRACE is a framework for efficient cross-embodiment adaptation of neural whole-body controllers, preserving pretrained motion representations while specializing feedback control to the target robot. It reduces embodiment mismatch through kinematic priors and aligns joint dynamics, then combines encoder distillation with low-rank reinforcement learning on a compact motion set. The method achieves 96.08% success on 67,953 evaluation motions with minimal GPU-hour cost, demonstrating practical reuse of a pretrained humanoid controller across robots with different body scales and dynamics.
Source: arXiv Robotics — research abstracts · Read original article ↗
Loading article text…
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
Reported numbers
motion frames
10,000,000
training hours
≈8 hours
evaluation motions
67,953
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
from Unitree G1 to Unitree H2 using 10 million motion frames and approximately eight hours of training on a single RTX PRO 6000 GPU, at less than 0.1% of the GPU-hour cost of training SONIC from scratch. The adapted controller achieves 96.08\% success on 67,953 evaluation motions under a height-scaled tracking criterion. Real-robot demonstrations span dynamic motion tracking and whole-body teleoperation, with collect
Open source S4
Source：arXiv Robotics — research abstracts · arxiv.org