Build AI datasets and physical AI reading

Egocentric-10K and Egocentric-100K are human-video datasets. The publisher cards specify Apache 2.0 and gated Hugging Face access. The 100K card specifies 456×256 video; do not assume the homepage’s 1080p specification applies to it. Check the 10K evaluation set and 100K evaluation set for the publisher’s evaluation materials; these are separate from the training releases.

Build AI’s website links Egocentric-1M to an X announcement and places larger releases on a future roadmap. This review did not confirm a downloadable 1M release or its license. Announced and planned data are not counted as available dataset listings. Do not add successive release sizes together without confirming whether recordings overlap.

SHOW3D is a separate Meta/Yale project linked in Build AI’s reading list, not a Build AI dataset. It provides synchronized human hand–object observations and annotations. Its published release uses CC BY-NC 4.0. Check the annotation version and scene-level availability before use.

Dyna-2 studies how human-video pretraining relates to robot learning. Read the report’s training sources, held-out tests and robot post-training conditions before interpreting its scaling claims. The paper is a research resource, not a downloadable corpus listing. GenCeption explores video-model pretraining for perception, including depth, camera pose and human understanding. This may help physical-AI perception work; it does not supply a robot-control dataset. These projects remain separate from Build AI despite being linked by its site.

For robot imitation learning, ask whether the release includes aligned robot actions. For learning from human video, inspect viewpoints, calibration, annotations and transfer assumptions. Use the dataset evidence checklist and compare the source-linked dataset directory. A reading-list link does not establish affiliation, data provenance or independent validation.