Egocentric Multi-View Dataset with 21-points 3D hand Keypoints Annotations

This dataset contains 10 multi-view egocentric household-activity demonstrations captured on head-mounted stereo camera and wrist cameras (no robot involved). Each demonstration is packaged as a HuggingFace LeRobot dataset (single episode per demonstration) at 1920 x 1456 landscape, 30 fps. Frame-level COCO-derived annotations are embedded as parquet columns alongside the video: 21-point 3D hand keypoints (left + right), hand visibility flags, top detected object class + confidence + bbox, segment-level action phase, and segment-level action verb. Using this dataset helps teams pretrain vision-language-action models on egocentric data, benchmark hand-object interaction perception, train action-recognition models, develop hand-keypoint regression heads, and bridge human-demonstration data into robot-learning pipelines.