Humanoid behavior foundation models (BFMs) enable diverse whole-body control, but adapting a pretrained controller to a new embodiment remains challenging because of kinematic and dynamic mismatch. We present TRACE, a framework for efficient cross-embodiment adaptation that preserves pretrained motion representations while specializing feedback control to the target robot.TRACE first injects kinematic priors into input and output projection layers and aligns nominal joint dynamics to reduce embodiment mismatch and support stable fine-tuning. It then combines encoder distillation with low-rank reinforcement learning on a compact, quality-screened motion set, using a two-phase curriculum to establish stable tracking before refining difficult motions. We apply TRACE to transfer SONIC 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 collected demonstrations supporting downstream vision-language-action policy post-training. These results demonstrate practical reuse of a pretrained humanoid controller across robots with different body scales and dynamics.
Figures & tables
Fig. 2: Overview of TRACE. Stage 1 (top left): projection layers are inserted at the reference, observation, and action interfaces to inject kinematic priors; action scales are calibrated and nominal joint responses are aligned. Stage 2 (bottom left): retargeting, quality screening, category balancing, and command-space selection form a compact motion dataset. Stage 3 (bottom right): with base weights frozen within each training phase, low-rank residual branches adapt the target Robot reference encoder through latent and token distillation, and the Robot Control Decoder through PPO in target simulation (Fig. 3 ). Stage 4 (top right): the transferred controller supports diverse robot tasks, including reference-motion tracking, VR teleoperation, and downstream VLA-based loco-manipulation.
Fig. 4: Convergence under a common motion-sampling distribution (difficulty-based oversampling disabled), with one run per method: (a) completion rate; (b) mean reward. TRACE converges within a small iteration budget; Direct Finetune requires substantially more training. Scratch fails to obtain a usable tracking policy within the 3k-iteration budget.
Fig. 5: Motion tracking across embodiments and deployment settings. From top to bottom, rows show Unitree G1 simulation, Unitree H2 simulation, and H2 hardware execution. Columns show (a) squatting, (b) single-knee kneeling, (c) kicking, and (d) single-leg hopping.
Method
Robot
SR (%) ↑
Eloc (mm) ↓
Evel↓
Eacc↓
SONIC (Pre-trained)
Unitree G1
97.59
27.81
5.157
1.390
TRACE (full)
Unitree H2
96.08
55.81
7.536
2.394
w/o-Pre-Adaption
Unitree H2
87.29
72.88
8.033
2.474
w/o-Data-Strategy
Unitree H2
93.80
54.01
7.853
2.558
w/o-Two-Phase-FT
Unitree H2
95.64
62.47
8.115
2.553
TABLE I: Results on the PHUMA benchmark. SR uses body-size-adjusted thresholds: 0.250 m (G1) and 0.3447 m (H2). Errors are evaluated only on each policy’s successful motions: Eloc in mm, Evel in mm/frame, and Eacc in mm/frame 2 , at 50 Hz.
Fig. 6: Data selection statistics. (a) Geometric mean of feature standard deviations and regularized covariance log-volume of frozen SONIC descriptors under shared standardization. Coordinates average 20 draws of 5,000 clips; bubble areas indicate clip counts. (b) Category frame shares; the dashed polygon marks equal shares (10%). H denotes normalized category entropy.
Fig. 7: Whole-body teleoperation and downstream applications on H2: (A) door opening and passage, (B) soccer, (C) toy packing, (D) trash disposal, and (E) chair repositioning. Keyframes progress left to right, showing execution across objects and contact conditions.
Metric
Single- phase
Two- phase
Δ (%)
G1 sim rollout
Training reward ↑
62.57
70.04
+11.9
—
Root-position error ↓
42.45
31.34
−26.2
18.62
Global foot error ↓
42.94
31.81
−25.9
19.04
Local hand error ↓
7.09
6.40
−9.7
6.31
Joint-angle error ↓
0.12555
0.11957
−4.8
0.1146
Critic value loss
0.28871
0.20108
−30.4
—
TABLE II: Final H2 SONIC transfer training statistics, with simulation rollouts using the pretrained G1 checkpoint under uniform sampling as a reference. Position errors: cm; joint-angle error: rad.
Task
Type
Trials
Success (%)
Object relocation (toy)
Single-arm
20
90
Object relocation (cube)
Bimanual
20
75
Object pickup (pillow)
Whole-body
20
80
Open door
Whole-body
20
65
Open trash can
Whole-body
20
65
TABLE III: VLA task success rates on the physical H2 robot.