Cross-Embodiment Transfer of a Frozen Humanoid Whole-Body Controller via Analytic Codec and LoRA Adapters

Sitarama Chekuri · Claude Fable 5

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

📄 Paper (draft PDF) 🤗 Models 💻 Code — run it (`./play_v2.sh`) ↩ Companion: frozen planner transfer SONIC-X2 port

A frozen, publicly released whole-body controller (GEAR-SONIC, trained on the Unitree G1) drives the AgiBot X2 Ultra — every encoder, the token bottleneck, and the motion prior untouched. A closed-form joint-space codec plus LoRA adapters on one decoder (0.25% of the platform's parameters), trained overnight on one 8-GPU node — about 2% of the platform's cited training compute.

The result is not parity but reversal: on out-of-distribution content the transfer beats the tracker natively trained for the X2 (69.0% vs 59.0% success) — while in-distribution benchmarks cannot tell the two apart. The same recipe, codec removed, specializes the platform on its own body with zero forgetting.

New in v0.2 (2026-09-12): the reversal survives hardware — an 18-minute untethered rough-terrain walk (grass, ramp, curb, pebbles; no e-stop, tilt p99 9.0°, three ≥20° excursions each recovered within 0.8 s) where the native line showed close-call imbalance on the same ground minutes later, running hotter. Plus: the platform's human-pose (SMPL) input path recovered to incumbent parity with a 0.3 sampling share, and the embodiment floor's actuator share anatomized — including a reward-margin pathology in which a 0.8 margin silently re-imposed a 19.2 N·m "hardware" ceiling.

How it works

System architecture: frozen GEAR-SONIC platform wrapped by the analytic codec, LoRA adapters on the dynamics decoder

Everything blue is the released platform, bit-identical and frozen. Everything orange is ours: a closed-form per-joint codec on the way in and out, and LoRA adapters riding the dynamics decoder — 0.25% of the platform's parameters, the only thing trained. The paper covers what the freeze makes measurable: the embodiment cost floor, the information bottleneck, and the overtraining arc.

Relaxed walk — incumbent vs zero-shot vs adapted

Left: the X2's natively trained incumbent tracker (~1,600 GPU-h curriculum). Middle: the frozen platform through the codec alone — zero gradient steps: geometry survives, visibly choppier, and this similarity-enabled survival is a precondition measurement, not a method claim. Right: + LoRA adapters (the selected transfer): the dynamic gap absorbed. All wrists held at the deploy-default pose. G1-stock reference column coming (pending a validated G1 MuJoCo recording harness).

Gangam dance — incumbent vs zero-shot vs adapted

Same three-way on an expressive dance. The adapted transfer tracks the choreography the incumbent renders conservatively; the zero-shot column shows the frozen prior's raw motion quality before any dynamics adaptation.

Results at a glance

model	novel500 (in-dist)	hard300v3 (tail)	PHUMA (OOD)	PHUMA survival
incumbent (native X2)	96.4 / 33.6 / 43.8	70.0 / 40.6 / 56.0	59.0 / 42.6 / 67.4	87.4
transfer (selected)	96.2 / 33.1 / 42.9	72.3 / 39.9 / 55.8	69.0 / 41.7 / 60.6	90.7

Success % / mean mm / p95 mm (survivors), IsaacLab strict gates, same machine; survival = MuJoCo second referee on the OOD slice. In-distribution: indistinguishable. Out-of-distribution: the frozen prior's edge, widened by polish.

Evaluation under the SONIC paper's criterion

Success rate (%)

set	stock G1 (own body)	39k incumbent (X2)	43k incumbent (X2)	16k frozen core (X2)	35k frozen core (X2)	40k frozen core (X2)
novel500 (in-distribution)	97.8	97.2	97.8	98.0	98.0	96.6
hard300v3 (in-distribution tail)	89.0	73.3	72.3	77.7	76.7	75.7
PHUMA 1,931 (out-of-distribution)	86.8	66.2	66.1	76.7	71.0	68.5
Pico captures via the SMPL encoder (27)	40.7	66.7	51.9	37.0	59.3	59.3

Tracking error (MPJPE-L, mm, successful clips only)

set	stock G1 (own body)	39k incumbent (X2)	43k incumbent (X2)	16k frozen core (X2)	35k frozen core (X2)	40k frozen core (X2)
novel500 (in-distribution)	26.9	34.3	33.6	31.9	32.7	32.4
hard300v3 (in-distribution tail)	35.3	41.5	39.9	38.6	39.4	39.0
PHUMA 1,931 (out-of-distribution)	32.8	43.4	43.7	41.6	41.3	41.3
Pico captures via the SMPL encoder (27)	40.4	42.4	46.1	51.5	43.2	42.4

Paper criterion (arXiv 2511.07820): a clip fails if the root or an end-effector height deviates by more than 0.25 m or the root orientation by more than 1 rad, fixed thresholds; MPJPE-L over the successful clips. Scored in IsaacLab, 256 parallel environments, on the same machine; the paper scores in MuJoCo, so its own numbers (PHUMA 97.2% on the full 68,326-motion set) are not directly comparable to our stock-G1 anchor. Sets: novel500 and hard300v3 are drawn from the BONES-SEED family the models were trained on (no explicit held-out split was used); PHUMA is the paper's only external benchmark and the only out-of-distribution row; the Pico captures are our own full-body VR teleop recordings, in the training corpus of the X2 models (in-distribution for them, unseen by the stock G1) and evaluated through the SMPL encoder, the live teleop input path. Headline checkpoints: the 39k incumbent (best on the Pico path) and the 35k frozen core (best X2 model on the general sets); the 16k frozen core is the deployed model and remains the strongest X2 checkpoint on PHUMA — the fine-tune toward the Pico captures lifted the Pico rows and cost points on the general sets.

Built on the released GEAR-SONIC platform. Companion work: frozen-planner transfer. Closest prior work: Any2Any (LimX Dynamics) — low-cost parity across dissimilar humanoids; this page: frozen-platform reversal on a closely-matched pair.