Physics-Based Motion Imitation
for Diverse Human Body Shapes

128 body shapes, one clip, one policy

Each video shows the same motion clip replayed across all 128 trained body shapes simultaneously, one shared policy controlling every shape at once, camera slowly orbiting the grid. These are trained-body demonstrations, not the new unseen-body panels. They illustrate simultaneous execution; they do not separately show the reference or measure sliding and penetration.

Abstract

Animating digital humans with different body proportions requires references consistent with their geometry and a controller that tracks them in physics. We study shared tracking, unseen-body transfer, and network-design trade-offs using one morphology-conditioned temporal controller. A paired corpus links 20,951 motion identities to body-conditioned references and matching simulation assets for 128 body configurations; training uses 18,855 identities across all bodies. The controller achieves 99.14% tracking success on 1,048 held-out motions, with one selected trained body per clip. On 233 training-split clips replayed across each of two 128-body unseen panels, success is 97.28% inside the training coefficient range and 97.15% outside it. These averages coexist with individual-body success as low as 3.0% and 13.3%, respectively. A later checkpoint improves outside-panel success but introduces substantial individual regressions. Three-seed comparisons of temporal multilayer perceptron and attention architectures reveal trade-offs between tracking and smoothness. An offline reference audit measures reduced stance-foot sliding and below-floor body centres with small pose changes, at the cost of increased reference jerk. These reference measurements are distinct from simulated tracking accuracy and smoothness; rollout sliding and penetration have not been quantified. The results demonstrate strong average shared tracking and transfer within the evaluated domain, while leaving all-body held-out-motion coverage, the causal contribution of explicit morphology inputs, and severe failure mechanisms unresolved.

Contributions and evaluation

1. Paired corpus and workflow: 20,951 clip identities on 128 body configurations (2,681,728 clip/body variants), each identity preserved from generation through simulation and replay.
2. Shared controller: one temporal slot/type-attention controller with streamed residency, trained directly on 18,855 clips across all 128 bodies.
3. Exposure-separated evaluation: 1,048 clip-disjoint test clips on trained bodies, and 233 training-split clips replayed on 128 unseen inside-range and 128 unseen outside-range bodies (29,824 clip/body pairs per panel), retaining every body and failure.
4. Design comparison and reference audit: three-seed comparisons of four network architectures, and a reference-refinement audit of 19,200 development references (150 clips × 128 bodies).

Tracking and unseen-body results

Success requires peak mean body-position error no greater than 0.5 m over the evaluated horizon. This is a tracking criterion, not a certification of visual realism. Unseen-body results were recorded on 9 October 2026.

Inside coefficient vectors lie in [−3, 3]¹⁰; outside vectors have max |β| in (3, 5]. Each panel has 64 vectors on two body templates. “Inside” refers to the coefficient box, not a demonstrated convex hull. The control uses different clips and a different refinement run, so its gap from unseen results is descriptive.

Average transfer is not uniform robustness. The lowest original-checkpoint body success is 3.0% inside and 13.3% outside. Outside-tier success is 99.03%, 97.76%, 95.33% and 96.47%, rather than a monotonic decline with coefficient distance.

Original: epoch 28,170; later: epoch 34,000. Training duration and motion exposure change together; this is a checkpoint comparison, not an isolated treatment. Confidence intervals resample clips while holding the body panel fixed. The unseen headline excludes 23 validation/test identities seen by the later checkpoint. A repeat of the original checkpoint changes headline success by at most 0.04 percentage points, but is not a training-seed replicate.

Network comparison

Each seed is summarized over its last eight evaluations at a matched 6,000-epoch budget on 150 training clips × 128 bodies. Position and rotation are across-seed means. Lower errors and jerk are better under the fixed evaluator convention; capacities differ. All networks receive morphology inputs. These supporting reruns do not establish statistical superiority or isolate the effect of explicit morphology conditioning.

Reference quality versus simulated motion

Mean reference changes are 0.12° in body rotation and 0.008 m in horizontal root position; the cross-shape knee-range variance ratio is 0.94. Refinement reduces the measured reference artifacts but increases reference jerk. Below-floor body centres are not mesh-penetration measurements. Raw reference jerk and the simulated normalized-jerk score have different definitions.

The held-out test measures simulated body-position error of 0.0676 m, rotation error of 0.1723 rad and normalized jerk of 1,050.9. Simulated sliding and penetration have not been quantified. All-body held-out-motion coverage and controlled reference-versus-simulation visual comparisons remain open.

Published capabilities and our contribution

These sources describe capabilities and evaluation scope, not directly comparable performance. No external tracker has been rerun on our matched references, assets and protocol.

Our additions are paired corpus assembly and identity checks, reference processing and its fidelity audit, the evaluated temporal controller configuration, streamed clip residency, and evaluation that separates motion exposure from body exposure. SMPL/SMPLSim, HUMOS and ProtoMotions provide the underlying body models, generation and learning infrastructure.

Pipeline

How one shared policy ends up controlling 128 different body shapes: matched data construction, then a training loop where the body shape, its actuation, and its reference all stay consistent per environment.

Dataset & Code

The full multi-morphology motion corpus is derived from AMASS via HumanML3D clip identities and HUMOS shape-conditioned generation. Derived motion tensors are not redistributed here; use remains subject to the source datasets' access and licensing conditions. The research code, the simulation body models and the evaluated checkpoint are linked below. Reproducing the evaluations requires the motion corpus, or references regenerated with the released pipeline.

• Code & pipeline: github.com/hansen1416/morph_mimic (version v1.0)
• Body-shape definitions & evaluated simulation assets: included in the code repository under protomotions/data/assets: the 128 trained bodies, the two 128-body unseen panels and their shape coefficients. The body models are derived from SMPL and released with attribution (CC BY 4.0).
• Trained checkpoint: huggingface.co/YugoAmaryl/morph-mimic (epoch 28,170, CC BY-NC 4.0). The epoch-34,000 checkpoint is not released.
• Motion data and evaluation archive: the motion corpus is not distributed; the released pipeline can regenerate references from the source datasets. No unrestricted derived-data release is claimed.

BibTeX

@misc{han2026morphmimic,
  title  = {MorphMimic: Physics-Based Motion Imitation for Diverse Human Body Shapes},
  author = {Han, Longzhen and Farrington-Wheeler, Beck and Hussain, Sajjad and
            Baimagambetov, Almas and Polatidis, Nikolaos},
  year   = {2026},
  note   = {Manuscript in preparation},
}