fastwalker1118/b1k-task5-navplace-2026-v1

BEHAVIOR-1K task 5 nav + place -- 2026 format

Training-ready LeRobot v3 data for BEHAVIOR-1K task 5, setting_mousetraps. This dataset intentionally contains only generated navigation and placement/recovery demonstrations. It contains no official human demonstrations and no generated full-task success episodes.

Contents

Source family	Episodes	Frames	Hours at 30 fps
Generated navigation	1,170	562,590	5.209
Generated placement / recovery	4,388	607,744	5.627
Total	5,558	1,170,334	10.836

• LeRobot codebase format: v3.0
• robot: R1Pro
• state: float32[61], BEHAVIOR 2026 convention
• action: float32[23], R1Pro VLA action
• RGB: three 224x224 H.264 streams using official 2026 camera keys
• frequency: 30 Hz

State and qvel convention

observation.state follows the compact 2026 layout:

Slice	Meaning
0:3	base-frame observed base_qvel = [vx, vy, wz]
3:10, 10:17	left arm qpos, qvel
17:20, 20:24	left end-effector position, quaternion
24:26, 26:28	left gripper qpos, qvel
28:35, 35:42	right arm qpos, qvel
42:45, 45:49	right end-effector position, quaternion
49:51, 51:53	right gripper qpos, qvel
53:57, 57:61	trunk qpos, qvel

The generated legacy sources carry world/canonical-frame holonomic qvel. The builder explicitly rotates x/y using the virtual-joint base yaw into the robot/base frame; it does not infer the source frame merely from state width. Observed qvel is not a normalized action and should not be expected to equal action[0:3] sample-for-sample.

Cameras

• observation.rgb.zed_link_camera_0
• observation.rgb.left_realsense_link_camera_0
• observation.rgb.right_realsense_link_camera_0

The generated sources do not contain depth or robot-to-camera poses. Those fields are omitted rather than fabricated.

Validation

• Manifest provenance: exactly 1,170 nav and 4,388 place episodes; no human, success, or unknown source entries.
• Every state is 61-D, every action is 23-D, and every value is finite.
• Episode, frame, global, and task indices are contiguous and agree with the metadata totals.
• All 16,674 output videos were decoded and matched to episode frame counts.
• Nav base-qvel standard deviation is [0.0669, 0.0345, 0.0358]; place is [0.0246, 0.00560, 0.00716].
• No episode has dead/zero-filled base qvel, including every episode with an active base action.
• The builder's sustained arm-qvel rail-corruption check passed every episode.

The conversion manifest is included at meta/conversion_manifest.jsonl. The dataset was produced by build_lerobot_2026.py in the associated BEHAVIOR-1K challenge codebase.

Compatibility

This is the 2026 format. A legacy 2025 pipeline expecting a 256-D state, world/canonical-frame base qvel, or observation.images.rgb.* camera keys needs an explicit format adapter.