fastwalker1118/b1k-task5-mousetraps

task-05 · VLA successes + mined nav demos (LeRobot v3.0)

Merged training set for setting_mousetraps: the short successful policy rollouts plus the navigation-recovery demonstrations mined from failed ones.

	episodes	hours	source
VLA successes (<6 min)	74	6.07	37 unique × 2
mined nav demos	2,340	10.42	585 unique × 4
total	2,414	16.49	622 unique

fps 30 · robot_type R1Pro · 1,780,548 frames · single task setting_mousetraps

Features

feature	shape	notes
observation.images.rgb.head	224×224×3	h264, yuv420p
observation.images.rgb.left_wrist	224×224×3
observation.images.rgb.right_wrist	224×224×3
observation.state	256	full R1Pro proprioception
action	23	base_vel[0:3], torso[3:7], L_arm[7:14], L_grip[14], R_arm[15:22], R_grip[22]

Depth streams and observation.cam_rel_poses are not included. π₀-style training consumes images + state + action only, and cam_rel_poses (3 cameras × pos+quat) is a forward-kinematics function of the 256-dim state. Both remain in by_instance/.

Everything is 224×224 because the mined demos were recorded at that resolution; the successes were downscaled from 720×720 (head) / 480×480 (wrist) to match.

Frames and units — read before training on velocities

All velocities in observation.state are exactly as the simulator reported them. No frame conversion is applied. This is deliberate: at eval the harness feeds the policy extract_state_from_proprio(obs["robot_r1::proprio"]), which slices base_qvel straight out of live proprioception. Training data must therefore carry the same quantity, or the policy learns to rely on a signal it never receives at deployment.

What that means concretely, for anyone reading these columns as physics:

• base_qvel (253:256) and joint_qvel[0,1] (84,85) are the holonomic base virtual-joint velocities. Their frame is the articulation root, a fixed scene anchor — for this scene rotated 75.028 deg from world, with origin offset [0.2431, 2.0411]. They are NOT world frame and NOT robot frame.
• robot_lin_vel (152:155) / robot_ang_vel (155:158) come from the root link and ARE world frame. So the two velocity blocks sit in frames 75 deg apart.
• All of these are instantaneous values sampled at the tick boundary, not the average over the tick. With physics at 120 Hz and actions at 30 Hz, a stiff position-controlled joint has largely settled by the sample instant: integrating base_qvel recovers ~48% of the distance actually travelled, and the arm blocks are worse. Position columns are the smooth, self-consistent ones.

If you need true robot-frame velocity, derive it from base_xy/base_yaw odometry (backward difference) rather than rotating these columns — but do it on BOTH the training data and the eval wrapper, or you reintroduce the skew.

quantity	frame	units
base_qvel, joint_qvel[0,1]	articulation root (scene-fixed)	raw (m/s, rad/s), instantaneous
robot_lin_vel, robot_ang_vel	world	raw, instantaneous
action[0:3] base velocity	base/body	normalized by [0.75, 0.75, 1.0], tick-average
observation.state base_qpos	articulation root	m, rad
robot_pos, robot_ori	world	m, rad
eef_*_pos	base-relative already	m

Note the observation and the action use different frames and different time semantics. That asymmetry is the benchmark's, not this dataset's.

The success episodes remain byte-identical to the challenge parquets they came from.

Note the model's own input is the 23-dim vector from extract_state_from_proprio (base_qvel 3 + trunk_qpos 4 + arm_left 7 + L_grip 1 + arm_right 7 + R_grip 1), not the 256-dim state stored here. Only dims 0-2 of that input are velocity; arm_*_qvel and trunk_qvel are never read by the policy.

Encoding

Every video is h264 / yuv420p (CRF 18), matching what info.json declares and what rollout_success uses. Both source halves needed correcting to get there: the mined demos were recorded as h264 yuv444p (High 4:4:4 — poor browser and NVDEC support), and the downscaled successes were first written as MPEG-4 Part 2. Frame counts were verified per file after re-encoding, so images stay aligned with actions.

Where the two halves come from

VLA successes — of the 498 evaluated rollouts, 66 reached q_score.final == 1.0. The 37 whose simulator_time was under 6 minutes are included. Note the outcome directory in by_instance/ is not a reliable success marker: early-terminating successes are filed under success/, but the 8 that ran to the 20344-step horizon are filed under failure/ despite q_score == 1.0. Success is decided only by logs/metrics/*.json.

Mined nav demos — recovery motions mined from failed rollouts by seeding the sim to the failure frame, retrieving a stance from the task-5 demo bank, and planning a tuck + forward-arc-RRT transit to it. Recorded with MOP_VLA_OFF=1, so each episode ends at the VLA handoff: these are motion demonstrations, and none has been shown to convert through a policy. 585 mined from 137 seeds over 15,612 attempts; median episode 15.2 s.

Duplication is physical

Episodes are repeated as separate episode_index values over identical frames, each with its own index range and its own video file — the same convention rollout_success uses for its natively-successful instances. This upweights nav demos roughly 1.7:1 against successes in a fixed-epoch run. It adds no new information: the 10.42 h of nav data is 2.6 h seen four times.

meta/episodes carries _source_kind, _source and _copy so the 622 unique episodes can be recovered without regenerating anything.

Layout note

One episode per video file (from_timestamp 0, to_timestamp length/30), rather than packing several episodes per file as rollout_success does. Readers locate frames by (chunk, file, from_timestamp, to_timestamp), which this satisfies.

Verified

Video frame count equals action row count for all 622 source episodes, re-checked after re-encoding; global index contiguous over 0..1,780,547; per-episode frame ranges contiguous; 2,414 video files present per camera; every video decodes as h264/yuv420p.