TRACE: Efficient Transfer Learning for Neural Whole-Body Control Across Humanoid Embodiments

Can a large pretrained humanoid controller, a so-called behavior foundation model, be carried over to a robot with a different size, joint layout and actuators cheaply, while keeping the motion interface that teleoperation and VLA models already use?

Behavior foundation models such as SONIC learn one controller for thousands of motions by tracking reference clips, and SONIC needed more than 100 million motion frames and about 21,000 GPU-hours on Unitree G1. Each such model is tied to the robot it was trained on: a new body changes the number and placement of joints (kinematic mismatch) and changes how the robot responds to the same command because masses, actuators and contacts differ (dynamic mismatch). Retraining per robot is expensive, and humanoid hardware is changing every year.

One line of work trains a single controller over many humanoids or randomized bodies at once, which raises training cost and still performs worse on robots outside the training set than a dedicated policy. Another adds residual or delta policies on top of a pretrained tracker, but only for the same robot. The closest prior method, Any2Any, aligned joints and fine-tuned dynamics-related modules with low-rank adapters (LoRA), but did not explicitly keep the motion latents aligned across robots and kept the source robot's motion sampling, which the authors say slows convergence.

TRACE works in four stages. First it folds a kinematic map between the two robots, fitted from paired retargeted poses with a Jacobian-based prior, into the controller's input and output layers, rescales actions and sets the target robot's PD gains so each joint has the same natural frequency and damping as on the source robot. Second it builds a compact 10 million frame training set from the BONES-SEED motion corpus: clips with joint-limit violations, collisions or foot sliding are removed, ten motion categories get equal budgets, and within each category clips that are spread out in SONIC's own feature space are preferred. Third, with base weights frozen, LoRA adapters (rank 32) are trained in two ways: the target robot's motion encoder is distilled to produce the same latent tokens as the frozen source encoder, while the control decoder is tuned with PPO reinforcement learning in simulation. Training runs in two phases, broad stable tracking first and hard motions second, with adapters merged in between.

Transfer from G1 to H2 took about eight hours on a single RTX PRO 6000 GPU with 12,288 parallel simulated environments, which the authors put at under 0.1% of SONIC's training GPU-hours (hardware not normalized). On 67,953 PHUMA benchmark motions held out from training, the H2 controller succeeded on 96.08% of clips under a height-scaled criterion, against 97.59% for the original SONIC on G1; local position error was 55.81 mm against 27.81 mm. Removing the pre-alignment stage cut success to 87.29%, removing the data-selection strategy to 93.80%, and a single-phase schedule to 95.64%. The two-phase schedule raised training reward by 11.9% and cut root-position error by 26.2%. Full fine-tuning converged much more slowly and training from scratch produced no usable tracker within 3,000 iterations. On the physical H2 the controller performed squats, single-knee kneeling, kicks and one-legged hops, and served as the low-level layer for VR teleoperation. Demonstrations collected that way were used to post-train NVIDIA's GR00T N1.7, which reached 65% to 90% success over 20 trials on each of five tasks, from 90% for single-arm toy relocation to 65% for opening a door and a foot-operated trash can.

This is a preprint and has not been peer reviewed. The authors say their test covers one robot pair (G1 to H2) and one pretrained model (SONIC), and that the method assumes a reusable critic network, which other training schemes may not provide. There is no head-to-head comparison with Any2Any or multi-robot controllers because, the authors say, compatible implementations were not available. In our reading, both robots come from the same maker, so the dynamics gap is smaller than between vendors; H2 tracking error is double G1's in absolute terms; hardware results are shown as demonstrations and VLA success rates on 20 trials, not as measured tracking accuracy; and the eight-hour figure excludes SONIC's original pretraining, which someone still has to pay for.