PointMotionBench eval kit — private working copy
Everything vla_foundry/eval/evaluate_pmb.py (vlp repo) needs to score HOT3D, DAVIS and WorldTrack, mirrored from the
working box on 2026-10-29. Private: the source videos are third-party data (HOT3D: Meta; DAVIS 2017; WorldTrack /
St4RTrack incl. Aria Digital Twin) under their own licenses — do not make this repo public or share it.
| path | what | from |
|---|---|---|
pointmotionbench/ |
Ai2 annotations, tracks, captions, index maps, reconstruction scripts | allenai/PointMotionBench |
hot3d_rgbs/ |
1,272 extracted Aria RGB MP4s, one per source clip (<clip>_rgb.mp4) |
pointmotionbench/hot3d/extract_rgbs.py over bop-benchmark/hot3d train_aria |
davis/DAVIS/ |
DAVIS 2017 trainval 480p (JPEGImages, Annotations, ImageSets) | davischallenge.org |
worldtrack/worldtrack_release/ |
WorldTrack splits we hold (adt_mini, ds_mini) | St4RTrack |
Use
huggingface-cli download meilongzhang/pointmotionbench-eval --repo-type dataset --local-dir $ROOT
uv run python -m scripts.run_pmb_eval --model <label> --config <run>/config.yaml --checkpoint <ckpt> \
--pmb-root $ROOT/pointmotionbench --hot3d-rgb-root $ROOT/hot3d_rgbs \
--davis-root $ROOT/davis/DAVIS --worldtrack-root $ROOT/worldtrack/worldtrack_release \
--shard-root <dir holding the run's mm1m test manifests> # load_run_config checks they exist
# Path pass-through was added to scripts/run_pmb_eval.py on 2026-10-29. On older checkouts call
# `python -m vla_foundry.eval.evaluate_pmb` directly with the same flags.
HOT3D under MolmoMotion's protocol (the external comparison) is documented in the vlp repo's EVAL_TABLES.md and
docs/EVAL_SPEC.md §3.3. Never use --limit on HOT3D: clip names group by source recording, so a prefix is biased.
- Downloads last month
- 23