fastwalker1118/b1k-task5-mixed-v1

BEHAVIOR-1K task 5 mixed — 2026 format

Training-ready LeRobot v3 data for BEHAVIOR-1K task 5, setting_mousetraps. This replaces the previous dataset at the same repository ID, whose 256-D / qvel conventions were not uniformly 2026-compatible.

Summary

Source family	Episodes	Frames	Hours at 30 fps
Official 2026 human demonstrations	200	2,039,222	18.882
Generated navigation	1,170	562,590	5.209
Generated success demonstrations	37	327,684	3.034
Generated placement / recovery	4,388	607,744	5.627
Total	5,795	3,537,240	32.752

• LeRobot codebase format: v3.0
• robot: R1Pro
• task: setting_mousetraps (local task_index=0)
• observation state: float32[61]
• action: float32[23]
• frequency: 30 Hz
• RGB: three 224×224 H.264 streams

Observation state

observation.state follows the compact BEHAVIOR 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 first three values are an observed simulator state in physical velocity units. They are not the normalized action and should not be expected to equal action[0:3] sample-for-sample.

For generated sources that retained the legacy 256-D raw proprioception, the builder extracts the documented state blocks and rotates the holonomic virtual joint velocity at legacy indices 253:256 into the moving robot/base frame using the virtual-joint yaw at index 246:

vx_base =  cos(yaw) * vx_world + sin(yaw) * vy_world
vy_base = -sin(yaw) * vx_world + cos(yaw) * vy_world
wz_base = wz_world

Official 61-D 2026 human states are preserved directly. Source qvel provenance is explicit during conversion; the builder does not infer a 256-D source's frame from its dimensionality alone.

Action

action is the 23-D R1Pro VLA action:

base[0:3], trunk[3:7], left_arm[7:14], left_gripper[14],
right_arm[15:22], right_gripper[22]

Actions are preserved from each source episode. In generated recovery HDF5s, the base action is a normalized, interval-derived achieved motion; it is not an instantaneous copy of observed base qvel.

Cameras

The dataset uses the official 2026 RGB feature keys:

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

Generated sources do not contain depth or robot-to-camera pose observations, so those fields are omitted rather than fabricated. The BEHAVIOR 2026 RGB π0.5 baseline consumes the fields present here.

Transition fields

Every row also contains next.reward, next.terminated, and next.truncated, plus standard LeRobot v3 timestamp, frame, episode, global, and task indices.

Validation

The uploaded artifact was validated before publication:

• all 5,795 Parquets and all 3,537,240 state/action rows were scanned;
• every state is 61-D, every action is 23-D, and all values are finite;
• episode/frame/global indices, timestamps, task mapping, and metadata totals agree;
• all 17,385 videos were decoded and checked against episode frame counts;
• exact source-to-output checks passed for human, navigation, success, and placement episodes;
• samples from every source family loaded through LeRobot 0.5.2 with all three RGB tensors.

The conversion manifest is available at meta/conversion_manifest.jsonl. The converter used to build this artifact is omnigibson/learning/embodiedClaw_recovery/scripts/export/build_lerobot_2026.py in the associated BEHAVIOR-1K challenge codebase.

Compatibility warning

This is the 2026 format. Do not train a legacy 2025 pipeline expecting a 256-D state, world/canonical-frame base qvel, or observation.images.rgb.* camera keys on this dataset without an explicit format adapter.