DeskTaskTac/RawHandPip/README.md

RawHandPip Sample

This directory is a complete-episode sample of the DeskTaskTac RawHand pipeline.

RawHand episodes use image-based hand keypoint detection and multi-view triangulation to recover hand state, together with tactile glove readings. This subset uses the same per-frame schema as ../AprilTagPip; see ../README.md for the DeskTaskTac field dictionary.

Summary

item	value
episodes	10
frames	10,974
task entries	10
top-view videos	10

Episode indices and video paths are local to this subset and align with the files under data/ and videos/video.top_view/.

Layout

• data/episode_*.parquet: frame-level episode data.
• videos/video.top_view/episode_*.mp4: top-view MP4 videos.
• meta/features.json: 38-field DeskTaskTac per-frame schema.
• meta/dataset_info.json: dataset-level metadata, modality mapping, and coordinate systems.
• meta/episodes.parquet: episode-level metadata. RawHand includes additional raw-hand processing fields.
• meta/tasks.jsonl: task id to task-name mapping.
• meta/episode_records/*.json: per-episode metadata records.

Loading Notes

• Resolve paths in video.top_view relative to this directory.
• Treat state.eef_* and state.left_eef_* as right/left hand wrist or anchor poses in the first-person OpenCV camera frame.
• timestamp is episode-relative seconds; timestamp_ms preserves the captured millisecond timeline.
• manus_frame_index is commonly -1 in RawHand episodes because this pipeline does not rely on a Manus/MoCap source in the same way as AprilTag episodes.