01Layer
L1–L5 Control Annotation
L1–L2 full corpus
The supervision a policy actually trains on — not the labels a video search engine needs.
Most video annotation answers what a clip contains. Control annotation answers what a policy has to reproduce: which phase the operator is in, which object is involved, which joints are touching it, and whether the attempt succeeded.
Control L1–L2 — task identity, outcome and phase boundaries — ship on every clip. L3 and above depend on reconstruction, hand pose and object tracking, so their coverage follows the published yield rather than the full corpus.
ANNOTATION DEPTH LADDER
L1 Task
Task identity and success/failure outcome.
L2 Phase
approach → grasp → transport → manipulate → release.
L3 Object
Open-vocabulary object identity, material and state transitions.
L4 Keyframe
1 fps dense keyframes plus every phase-boundary frame.
L5 Contact (inferred)
Per-hand, per-joint contact inferred from MANO surface distance, derived from MANO surface distance.
SPECIFICATION
Coverage
Control L1–L2 across the full corpus · L3–L5 per published yield
Review
Expert adjudication on a sampled subset
Agreement
Published per delivery on the data card
Available on MOVAS-Ego 100K. Coverage per scene family is published on each family page and on the data card.