Zero-Shot Cross-Embodiment Reuse of a Frozen Humanoid Motion Planner

Sitarama Chekuri · Patrick Rose · Claude Fable 5

Ultimate Bots · Anthropic (AI co-author) · draft preprint, 2026

📄 Paper (draft PDF) 🤗 Models (SONIC-X2) 💻 Code (release pending) 🕹 Web demo — X2 & G1 (Kimodo) ↩ SONIC-X2 port (companion report)

A 2M-step motion planner trained entirely on the Unitree G1 drives the AgiBot X2 Ultra at deployment quality — with zero gradient steps on its weights. An analytic joint-space wrap plus an 83k-parameter per-frame residual head, selected purely by closed-loop rollouts, beats both planners natively trained for the X2.

~10,000 H100-hours (the source planner's training schedule) replaced by ~6 hours on one workstation GPU of adaptation. The central finding is methodological: offline fit and closed-loop viability diverge — every time an offline metric argued, it argued wrong.

Slow walk, 0.3 m/s — the demo gait

Source robot (left, native planner) vs target robot (right, wrapped planner + heads + walk-start templates), identical velocity commands. Kinematic reference playback, daemon-faithful 16-frame serving. Watch the first step: the source's native gait initiation surges — a trait of its training corpus — while the target's first stride matches its steady gait, courtesy of operator-recorded initiation templates injected into the frozen planner's clip library (Section 9 of the paper). Operator verdict after driving it in simulation: "ready to ship."

Run, 3.0 m/s command — stress case

Out of the deployable envelope — bounded by the source core's own corpus and the tracker's training coverage, not by the transfer (an A/B isolation in the paper shows the native source core fails the same speed asks alone). Stride style transfers faithfully.

Crouch walk — a capability neither native planner had

Inherited for free from the source clip library through the wrap; commanded depth saturates at the source corpus's floor.

Headline numbers

closed-loop metric (walk 0.35, sim)	old native planner	new native (32-GPU run)	borrowed core (ours)
recovery steps, seam sweep	6	18–36	{1,2,1}
operator tape A (350 s, 84 stops)	58 extra steps	68, fell @238 s	26, full survival
operator tape B (103 s, 16 stops)	34	24	6
stance-height wobble	±0.2 cm	—	±0.4 cm
new-model training cost	~8×H100, weeks-class	~1,500 H100-h	~6 RTX-5090-h

Cloud training for the native attempts and the SONIC-X2 tracker ran on Nebius-sponsored GPU credits (4×8 H100, InfiniBand). The source planner is MotionBricks (NVIDIA), used as released, frozen. Tracker checkpoints, training configs, and continue-training notes: huggingface.co/tinkerbuggy/sonic-x2.

Draft page — links and the PDF will update as the preprint finalizes.