Integration view
Modality × annotation matrix
For each sensor modality: what gets labelled, how, with which tools, in which storage format and who consumes it. Source: annotation_matrix in data/stage-04.js. Click the Modality header to sort.
Two patterns to notice
Dense visual labels (masks, poses, captions, segments) are now mostly model-generated (SAM 3.1, VLMs, FoundationPose) with human audits. Contact modalities (F/T, tactile, audio) are labelled as events with timestamps rather than dense annotations, and they often provide the most reliable success evidence for contact-rich tasks. See stage 04.