fastwalker1118/b1k-task5-nav-place-raw

task-05 setting_mousetraps — nav demos + generated place demos (raw bundles)

Raw per-episode artifacts for two sources, uploaded together for inspection and debugging. This is not a training-ready dataset — it is the material each source is built from, in each source's own native format. The two halves do not share a format; see below.

	episodes	format	size
nav/	615 bundles (585 unique)	hdf5 + composite mp4	~5.6 GB
place/	2,194 clips over 200 source episodes	parquet + 3x per-camera mp4	~0.8 GB

nav/ — navigation-recovery demonstrations

Mined from failed policy rollouts: the robot is restored to a judged failure frame and a recovery is attempted; the successful ones are kept.

nav/<run>/<bundle>/recovery_vla.hdf5     phase, step_idx, task_index, stage,
                                         observation, action, base_xy, base_yaw
                  /rgb_composite.mp4     672x448 3-camera composite
                                         (wrists stacked left, head right)
                  /contact_trace.json
nav/<run>/_method_result.json            run outcome
nav/<run>/sequential_report.jsonl        per-attempt log

615 bundles cover the 585 distinct episodes that reached the LeRobot build, plus 30 extra. Run/bundle names encode the source instance, seed and frame, e.g. navdemo_idx304_s20343_1438/navdemo_51840_6720_left.

place/ — generated noise-injected place demonstrations

Each clip is a human demo's pair of mousetrap releases — 1 s lead-in with both traps held, release A, the gap, release B, 1 s tail — replayed from a constant base offset with no correction motion, kept only if both traps satisfy their BDDL atoms and the starting pose is collision-free.

place/ep<SOURCE_EPISODE>/2025-challenge-demos/
    data/task-0005/episode_<demo_id>.parquet          observation.state (256),
                                                      action (23), cam_rel_poses (21)
    videos/task-0005/observation.images.rgb.{head,left_wrist,right_wrist}/
                     episode_<demo_id>.mp4            224x224 h264
    meta/episodes/task-0005/episode_<demo_id>.json
place/ep<SOURCE_EPISODE>/_ledger.jsonl                ONE ROW PER ATTEMPT
place/ep<SOURCE_EPISODE>/_summary.json
place/pairs.jsonl                                     the 384-pair manifest

_ledger.jsonl is the useful part for debugging. Every attempt, kept or rejected, records the noise drawn (eps_xy, eps_yaw), the outcome (ok / collision / bad_grasp), the per-trap BDDL atoms, how far each trap landed from the demo's spot, the gripper-to-trap distance, and how far the constant offset drifted.

demo_id encodes provenance: source episode tail, release frame, arm, attempt.

Generation summary (200 source episodes, 3,840 attempts)

kept	2,194 clips = 2.81 h, mean 4.62 s
unperturbed control pass	363/384
noisy kept	2,194/3,456 = 63%
rejected: pose in collision	741
rejected: object not at the gripper	40
offset drift over a clip	p50 0.0000 m, p90 0.0032, max 0.026
gripper-to-trap distance	p50 0.049 m (a held object sits 0.03-0.06 m from the eef link)
sink / toilet	956 / 1,232 clips

Noise is a constant rigid base offset held for the whole clip, sampled stratified over a 3x3 grid of a rectangle in world axes (the sink's yaw is 0 in every instance, so world x is its depth axis and world y runs along the basin):

anchor	depth	lateral	yaw
sink	±0.05 m	[−0.30, +0.20] m	±6°
toilet	±0.15 m	±0.15 m	±10°

Known caveat on observation.state

place/ carries the 256-dim 2025-convention proprioception, because it was recorded against that convention. The 2026 challenge release uses a 61-dim state, and the two disagree on every velocity channel — in the 2025 convention base_qvel[253:256] is zero-filled, and base linear velocity lives in robot_lin_vel[152:155] in the world frame rather than the base frame. Treat the velocity fields here as unreliable; positions are sound.

What is not here

No hdf5 for place/ — the generator wrote parquet + mp4 only, and the sim state needed to produce one was not retained per clip.