A foundation for the next embodiment

TRACE

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

Qingyu Lan1, Zhihao Tian1, Zifan Wang1, Junwei Liang1,2,†

1 The Hong Kong University of Science and Technology (Guangzhou)

2 The Hong Kong University of Science and Technology

† Corresponding author

Read the paper Live demo Video Github

TRANSFER DIRECTIONUnitree G1 Unitree H2

Explore the research

Reuse the knowledge. Adapt the control.

New body.
Same motion language.

Humanoid behavior foundation models learn a rich repertoire of movement. TRACE brings that repertoire to a different robot by preserving pretrained motion representations and adapting feedback control to a new body.

~8h

Single-GPU adaptation

G1 → H2 · RTX PRO 6000
96.08%

Motion-tracking success

67,953 PHUMA motions · simulation
10M

Curated motion frames

Compact, quality-screened transfer set

Success uses a height-scaled tracking threshold. Adaptation cost excludes source pretraining; GPU-hours are not normalized across hardware. Evaluation details ↗

Interactive playground

Meet the controller.

Explore whole-body motion tracking
and planner-controlled locomotion.

Unitree H2 · Interactive 3D environmentOpen full demo

The project overview

See TRACE in action.

From cross-embodiment adaptation
to whole-body skills on hardware.

Project overview · 2:45 · 1080p

The approach

Transfer more.
Train less.

Align the embodiment, curate the motions,
then specialize with low-rank adaptation.

TL;DR

Keep the pretrained motion language. Align the robot interfaces, distill compatible command tokens, and fine-tune the controller for the target dynamics.

TRACE framework: cross-embodiment pre-adaptation, motion dataset curation, parameter-efficient fine-tuning, and deployment on Unitree H2
The TRACE pipeline. A shared command interface supports motion tracking, VR teleoperation, and downstream VLA policies.
01

Align the embodiment

Inject kinematic priors into projection layers. Align action scales and nominal joint dynamics before training.

02

Curate the motions

Retarget and screen motion quality, balance categories, and select diverse commands within a 10M-frame budget.

03

Adapt efficiently

Distill reference-encoder latents and tokens. Use LoRA and a two-phase PPO curriculum to specialize feedback control.

04

Deploy across tasks

Reuse the adapted controller for dynamic tracking, whole-body teleoperation, and VLA-driven interaction.

On the real robot

One controller.
A repertoire of skills.

Large posture changes, dynamic balance,
and contact-rich whole-body interaction.

Hardware · 1× speed

Real-robot footage from the project’s demonstration materials. Individual demonstrations illustrate capabilities; quantitative evaluations are reported separately below.

Measured transfer

Broad capability.
A small adaptation budget.

Evaluated on held-out motions
and downstream hardware tasks.

Motion-tracking success

PHUMA · 67,953 motions
SONIC G1 · source97.59%
TRACE H2 · transferred96.08%
H2 component ablations
Without pre-adaptation87.29%
Without data strategy93.80%
Without two-phase fine-tuning95.64%

Simulation height-error thresholds: 0.250 m for G1 and 0.3447 m for H2. Each motion must stay within the threshold at every evaluated frame. Success rates use body-size-adjusted tolerances.

Downstream VLA tasks

Physical H2 · 20 trials per task
Task Control Success
Relocate a toy Single-arm 90%
Relocate a cube Bimanual 75%
Pick up a pillow Whole-body 80%
Open a door Whole-body 65%
Open a trash can Whole-body 65%

GR00T-N1.7 post-training uses demonstrations collected through the transferred whole-body controller. Tasks share the H2 low-level controller and motion interface.

Read the complete evaluation
Training efficiency & evaluation details
Completion rate and mean reward over training iterations for TRACE, Direct Finetune, and Scratch

Stable tracking, then difficult motions.

TRACE uses a two-phase curriculum: establish broad tracking first, merge and reset the LoRA adapters, then emphasize difficult motions. The reported endpoint comparison reduces root-position error by 26.2% relative to single-phase adaptation.

Training curves show one run per method under a common motion-sampling distribution, with difficulty-based oversampling disabled. The approximately eight GPU-hour adaptation excludes SONIC pretraining; hardware differs from SONIC’s reported 21,000 GPU-hour reference.

Build on a learned foundation

Whole-body intelligence,
ready for a different body.

Explore the paper Github

Citation

@misc{lan2026traceefficienttransferlearning,
  title={TRACE: Efficient Transfer Learning for Neural Whole-Body Control Across Humanoid Embodiments},
  author={Qingyu Lan and Zhihao Tian and Zifan Wang and Junwei Liang},
  year={2026},
  eprint={2610.17999},
  archivePrefix={arXiv},
  primaryClass={cs.RO},
  url={https://arxiv.org/abs/2610.17999},
}