Hyphenbox Turns Egocentric Video Into Robotics Training Data

Field notes

The homepage frames annotation as the layer between capture and model training, rather than presenting another robotics model. Its proposed pipeline combines frame-level action segmentation with 3D reconstruction: 21 keypoints per hand, 52 body joints, root trajectory, object state, and contact sequences. The material is aimed at robotics labs and collection partners handling first-person footage, with access apparently initiated through a consultation or research-team email.