Adapt inference
Create your own repository with the core as a root-level submodule. Implement reset, act and close. Preserve the author's deployment behavior.
SDK and adapter guide ↗CONTRIBUTE TO THE BENCHMARK
Make your result reproducible.
A model repository, a pinned checkpoint, and a verifiable run under the selected benchmark’s rules.
Run the whole suite
Use one checkpoint and the fixed protocol. Keep outputs outside Git. Submit full pose traces, hashes, receipts and inference-parity evidence.
Read the evaluation protocol →Request review
Open an issue linking your adapter and external result artifacts. Include the actual submitter, source/checkpoint versions, training-data overlap and any adaptations.
Open a submission issue ↗A small adapter. A shared kernel.
MotionChallenge-your-model/
├── MotionChallenge-Core/ # pinned core submodule
├── adapter.py # inference only
├── runtime.json # clocks, artifacts, input contract
├── provenance.json # source and review evidence
├── evaluate.py # shared core entrypoint
├── pyproject.toml
└── uv.lockThe core owns datasets, physics, metrics and verification. Model repositories do not copy or override evaluation logic. Checkpoints, data and results live in local caches or external artifact storage.
Propose another dataset
Publish a manifest with immutable source revisions, file hashes, frame rate, joint order and the complete motion set. Define a distinct protocol version and benchmark identity. Preserve difficult clips, document licensing and overlap, and keep scores in their own track.
Dataset integration contract ↗