MorphMimic: Morphology-Conditioned World Model and Policy Sharing Across Humanoids

Yang, Lizhi; Xie, Fengze; Wang, Yuyang; Gan, Lu; Yue, Yisong; Ames, Aaron D.

MorphMimic

Morphology-Conditioned World Model and Policy Sharing Across Humanoids

Lizhi Yang¹, Fengze Xie^1*, Yuyang Wang^1*, Lu Gan², Yisong Yue¹, Aaron D. Ames¹

¹California Institute of Technology ²Georgia Institute of Technology

* Equal contribution

Paper Code Video Demo

One World, One Model

Watch the film

The Unitree G1, AgiBot X2 Ultra and Booster T2 (left to right) dance a street shuffle with one shared policy. Simulated in MuJoCo with unified_v2 at iteration 76,000.

See it in your browser

The shared policy runs live in your browser (about 90 MB on first load). Drag a robot to push it.

Method

MorphMimic builds on an independent reimplementation of SuperMimic, which trains a tracker by differentiating through short rollouts of a learned dynamics model. A canonical layout maps each robot's joints by kinematic function onto a common joint order and 14 tracked bodies, and a 125-value descriptor of actuation, geometry and mass conditions the shared networks. Normalization statistics, reference motions and replay data stay robot-specific.

Diagram of the shared policy: per-robot reference goals pass through a shared motion encoder to 32-dimensional latents, a shared policy decoder with FiLM on the morphology vector, per-robot action maps and replay buffers, a five-point encoder below, and a shared world model for 32-step imagined rollouts. — **Figure 1.** Shared policy and command interfaces. Robot-specific normalizers feed a shared motion encoder, whose 32-dimensional latent enters the shared decoder with proprioception and morphology, and morphology also modulates decoder features through FiLM. Motion latents of corresponding frames align to their mean, and the five-point encoder matches its robot's motion latent.

Results

We compare four configurations, each trained for 40,000 iterations with 4096 environments per robot: a specialist per robot, a shared world model, a shared policy and critics, and the unified model that shares all three. Each of the 128,786 common clips runs once from its first frame under the deterministic policy. A shared configuration matches a specialist when its mean normalized return is within 0.02, a tolerance set before the evaluation.

Interface	Configuration	G1		X2 Ultra		T2
Interface	Configuration	Completion (%) ↑	Mean return ↑	Completion (%) ↑	Mean return ↑	Completion (%) ↑	Mean return ↑
Full motion	Specialist	99.797	0.9306	99.818	0.9253	99.860	0.9448
	Shared WM	99.734	0.9274	99.845	0.9269	99.861	0.9401
	Shared policy	99.806	0.9317	99.861	0.9278	99.775	0.9404
	Unified	99.747	0.9358	99.891	0.9304	99.886	0.9398
Five-point	Specialist	99.849	0.8807	99.841	0.8689	99.857	0.9014
	Shared WM	99.720	0.8681	99.851	0.8788	99.875	0.8979
	Shared policy	99.769	0.8781	99.847	0.8855	99.766	0.8976
	Unified	99.729	0.8626	99.880	0.8774	99.852	0.8890

Table I. The four 40,000-iteration sharing configurations on the 128,786-clip common set. Specialist has a separate world model and policy per robot, shared WM shares only the world model, shared policy shares the policy and critics, and unified shares the world model, policy and critics. Completion is the fraction of clips that reach the final frame without early termination. Higher is better (↑); the best value in each robot and interface block is in bold.

Every shared configuration stays within 0.02 mean return of its specialist on all six robot and interface pairs. The largest gap is 0.018, for the unified model on G1 five-point tracking. No configuration completes the most clips everywhere: the unified model leads on both X2 interfaces and on T2 full-motion tracking. The unified run uses 16.3 GPU-hours, against 13.9 for the three specialists together, and gives one set of weights for all three robots.

Across the four configurations, five-point commands lower mean return by 0.04 to 0.07 on each robot.

Five-point tracking: the G1, X2 Ultra and T2 walking and waving, with colored spheres at the pelvis, ankle and wrist targets, fading trails of past targets and curves for the ten-frame preview. — **Figure 2.** Five-point tracking with the shared policy during the walking (left) and waving (right) phases of one reference motion. Spheres mark the current pelvis, ankle and wrist targets, solid curves the ten-frame preview, and fading trails the past targets (display only).

Broad jump, jogging and a dance pirouette, five poses each, for the Unitree G1, AgiBot X2 Ultra and Booster T2. — **Figure 3.** Broad jump, jogging and dance (pirouette) with the same shared full-motion policy on the G1, X2 Ultra and T2. Each strip shows five simulator poses in time order from left to right. Headings and horizontal spacing are adjusted for display.

Ten clips in simulation

The G1, X2 Ultra and T2 (left to right) track each clip through the full-motion interface. Simulated in MuJoCo with unified_v2 at iteration 76,000. The results above are the paper's, at 40,000 iterations.

Walk

Ballet: pas de bourrée

Ballet: turns

Street: shuffle

Street: hip hop

Jive

Macarena

Country line dance

Dance routine

Robot dance

Hardware

We run the shared policy from Table I, trained for 40,000 iterations in simulation, directly on a physical Unitree G1, with feedback from its pelvis IMU and 29 joint encoders. Through the full joint-space motion interface, the G1 dances, walks and follows an operator who wears a PICO headset, two hand controllers and two ankle trackers.

Unitree G1 hardware: upper-body motion (full-motion interface)

Unitree G1 hardware: walking (full-motion interface)

Unitree G1 hardware: teleoperation (full-motion interface)

AgiBot X2 Ultra: hardware pending Booster T2: hardware pending

BibTeX

@misc{yang2026morphmimic,
  title        = {{MorphMimic}: Morphology-Conditioned World Model and Policy Sharing Across Humanoids},
  author       = {Yang, Lizhi and Xie, Fengze and Wang, Yuyang and Gan, Lu and Yue, Yisong and Ames, Aaron D.},
  year         = {2026},
  note         = {Technical report},
  howpublished = {\url{https://lzyang2000.github.io/morphmimic/}}
}