gimai/wnf-robot2human

WNF robot-to-human paired videos

Paired videos of the 12 Watch-N-Follow tasks: each robot episode and the same episode converted to a human hand. Each pair is frame-aligned (same frame count, 768x768, 30 fps).

Layout

<task>/
  robot_videos/ep<N>.mp4   robot episode
  hand_videos/ep<N>.mp4    the same episode with a human hand
  pairs.json               the pairing and per-episode metadata

pairs.json holds source_dataset and a pairs list. Each entry has episode, robot_video, hand_video (paths relative to the task folder), n_frames (30 fps), n_steps (10 fps source steps), source_episode and the two cameras.

Tasks

task	pairs	source	MB
InsertPeg	96	RoboMME (LeRobot format)	283
MoveCube	100	RoboMME (LeRobot format)	236
PatternLock	100	RoboMME (LeRobot format)	132
PickXtimes	100	RoboMME (LeRobot format)	625
RouteStick	100	RoboMME (LeRobot format)	273
StopCube	100	RoboMME (LeRobot format)	246
VideoPlaceOrder	100	RoboMME (LeRobot format)	1018
battery_try	48	RMBench	221
blocks_ranking_try	50	RMBench	402
observe_and_pickup	48	RMBench	23
put_back_block	200	RMBench	405
reopen_drawer	125	MemoryBench	368

Cameras

The two videos of a pair do not share a viewpoint. hand_camera is the human ego camera; robot_camera is the pipeline's robot view camera. Neither is the head camera of the source datasets. Both cameras are listed per episode in pairs.json.

Generation

10 fps source steps are replayed, the robot is replaced by a posed MANO hand with an SMPL-H arm, rendered and composited, then interpolated to 30 fps. Robot videos are the robot panel of the side-by-side output, re-encoded with x264 (CRF 18).

Excluded episodes

• InsertPeg: 1029, 1037, 1069, 1081 (hand contact limits not met)
• observe_and_pickup: 25, 35
• battery_try: 31, 46 (no usable source replay)

Known issues

• InsertPeg 1048 and 1090: a third peg leaves the table in the first frames; InsertPeg 1001: it drops onto the table. This is in the source demonstrations.
• InsertPeg contact loss above 10% (fingers lose the peg in more than 10% of held steps): 1045 (72.9%), 1039 (15.7%), 1096 (11.8%). The other 93 InsertPeg pairs are below 10%.

Which part of the source episode

RoboMME episodes have two phases: the robot first performs the variant once (is_demo rows, before exec_start_idx), the scene is reset, and the robot performs it again (execution phase).

task	clips cover	source_rows
InsertPeg, MoveCube, PatternLock, RouteStick, VideoPlaceOrder	the demonstration phase only	[0, exec_start_idx)
PickXtimes, StopCube	the whole episode (no demonstration phase, exec_start_idx = 0)	[0, length)
battery_try, blocks_ranking_try, observe_and_pickup, put_back_block (RMBench), reopen_drawer (MemoryBench)	the whole recorded episode, time-resampled	not given

For RoboMME tasks every pair carries source_rows, source_exec_start_idx and source_length (LeRobot rows, 10 Hz). Prompt videos made from the execution phase therefore do not match these clips frame for frame; the scene layout is the same, the robot motion and its length are not. In every pair n_frames = 3 * n_steps - 5.

Licence

Not yet decided. Access is gated; requests are reviewed manually until the terms of the source datasets and of MANO, SMPL-H and CoppeliaSim Edu are reviewed.