YugoAmaryl/morph-mimic

MorphMimic, epoch 28,170

One policy that tracks motion references on 128 SMPL body shapes in physics simulation (Isaac Gym). This is the checkpoint evaluated in the paper "MorphMimic: Physics-Based Motion Imitation for Diverse Human Body Shapes" (Han, Farrington-Wheeler, Hussain, Mubarak, Baimagambetov, Polatidis; University of Brighton). Paper: coming soon (arXiv).

• Code: https://github.com/hansen1416/morph_mimic
• Project page: https://hansen1416.github.io/morph_mimic/

Files

File	Contents
last.ckpt	The checkpoint at epoch 28,170 (Lightning checkpoint; PPO actor and critic). SHA-256 4b83f20c93bf95fab389fc8935685ef8a2c7e46583f765e8c802421edb4479cb
resolved_configs_inference.pt	The frozen inference configuration (pickle; the code needs it to rebuild the policy)
resolved_configs_inference.yaml	The same configuration in readable form
SHA256SUMS	Checksums of the three files above

The .ckpt and .pt files are Python pickles. Load them only from this repository, and only inside the code repository, which defines the classes they refer to.

Model

Temporal slot/type attention policy. Three observation encoders (current state, history, future reference targets) produce 12 tokens, which a two-layer, four-head Transformer (width 256) combines with learned temporal-slot and token-type embeddings. The 11-dimensional morphology vector (gender and ten shape coefficients) and the previous action bypass the attention and enter the separate actor and critic heads. The actor has six 2,896-unit layers and outputs 69 joint targets for a PD controller at 30 Hz; the critic has four 1,024-unit layers. About 51 M parameters in total. Trained with PPO, one seed, on six A40 GPUs for 28,170 epochs (about 11 billion simulation frames, roughly 450 GPU-hours).

Training data

References come from HUMOS on HumanML3D clips, generated for 128 SMPL body shapes (64 shape vectors on a male and a female template) and refined to reduce foot sliding and ground penetration. The motions derive from AMASS and are not distributed. The 18,855 training clips of the paper's split are used with a priority-based clip sampler. See the code repository for the pipeline.

Results (from the paper)

Held-out clips on trained bodies, one body per clip, 1,048 clips per split:

Metric	Validation	Test
Success at 0.5 m (%)	98.47	99.14
Success at 0.3 m (%)	95.52	96.28
Success at 0.2 m (%)	84.83	85.97
Success at 0.1 m (%)	37.88	38.55
Mean body-position error (m)	0.0699	0.0676

Unseen bodies (233 training clips replayed on 128 new bodies per panel): 97.28% success inside the training range of shape coefficients and 97.15% outside it.

Success means the time-maximum of the mean body-position error stays at or below the threshold. The validation split was monitored during training; the test split was evaluated once. The 0.5 m threshold is permissive, and the stricter thresholds partly measure the quality of the generated references. Success is high on average but concentrates failures on individual bodies (one unseen body succeeds on 3.0% of clips).

Intended use and limits

For research on physics-based motion imitation and body-shape generalization, in simulation. It is not validated on real robots or people, the controller's torques and powers are not calibrated biomechanics, and it needs the code repository and the (undistributed) motion corpus or your own references to run the evaluations.

Using the checkpoint

Evaluate with the code repository's evaluators, for example:

python protomotions/evaluate_validation_split.py --checkpoint last.ckpt --simulator isaacgym \
    --num-envs 128 --clip-batch-size 128 \
    --overrides motion_lib.r2_source=/path/to/hhi_stage2_per_clip_refined/

The stored configuration points at the clip source used in training, which is not public; the override replaces it. The same applies to resolved_configs_inference.pt.

License

Creative Commons Attribution-NonCommercial 4.0 International for the weights and configuration (the code is Apache-2.0). The training motions derive from AMASS, and the body models from SMPL, both licensed for non-commercial research.

Citation

The paper reference will be added when the arXiv version is available.

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