yeeeiii111/wuji-pick-and-place

Wuji Pick-and-Place Bundle

A paired human demonstration ↔ robot teleoperation dataset for bimanual pick-and-place. Built to train human-to-robot transfer policies (e.g. VAM-style action diffusion conditioned on first-person human reference video).

Both subsets are in LeRobot v2.1 format and share the same 8 in-domain tasks (task_index ∈ {0..7}). They live as separate sub-directories because the camera layout and resolution differ.

.
├── teleop/      LeRobot v2.1 — 179 ep, robot bimanual, 3 cameras (head + L/R wrist), 720×1280 head + 480×848 wrist
├── ego_ref/     LeRobot v2.1 — 179 ep, human egocentric, 1 cam (head), 720×1280
├── eval_refs/   LeRobot v2.1 — 102 ep, evaluation-only ego refs (8 in-domain leftover + 4 OOD), 1 cam (head), 720×1280
└── README.md

Sample counts per task

task_index	sub-task	top-level	teleop ep	ego_ref ep
0	ball_left	ball	30	30
1	ball_right	ball	30	30
2	cube_left	cube	30	30
3	cube_right	cube	30	30
4	cup_left_body	cup	15	15
5	cup_left_handle	cup	14	14
6	cup_right_body	cup	15	15
7	cup_right_handle	cup	15	15
total			179	179

Note: "cube" is physically a sponge block (instructions in tasks.jsonl say so). "cup body" vs "cup handle" denote which part of the cup the gripper grasps.

teleop/

• Format: LeRobot v2.1, fps 30
• Episodes: 179 (continuous episode_index 0..178, sorted by task_index then time)
• Frames: 75,783
• Robot: bimanual arms + dexterous hands (54-D state/action)
• Cameras (libx264 + yuv420p):
  • observation.images.head — 720×1280 (robot head-mounted)
  • observation.images.cam_left_wrist — 480×848
  • observation.images.cam_right_wrist — 480×848
• State: absolute joint positions read back from encoders, synced to the head camera @ 30 Hz (nearest-neighbor, max gap 25 ms). Layout: left_arm[7] + right_arm[7] + left_hand[20] + right_hand[20].
• Action: absolute target joint positions from /{left,right}_{arm,hand}/joint_commands, same layout. Arm joints converted to radians (source mcap stored them in degrees); hand joints stay in radians.
• Task prompt template: "Pick up the {object} {by the body|by the handle}? with the {left|right} hand and place it in the basket."

ego_ref/

• Format: LeRobot v2.1, fps 30
• Episodes: 179 (1-to-1 paired with teleop by task_index)
• Frames: 21,543
• Camera: observation.images.head (720×1280, h264 + yuv420p) — head-mounted GoPro, first-person view of a human performing the same task
• State / Action: 54-D zero-vector placeholders (no joint data for human demos)
• Task strings: same as teleop/tasks.jsonl

This subset is the reference video for VAM-style training. Pair by task_index: the i-th teleop episode within a task_index is paired with the i-th ego_ref episode within the same task_index (sort-order pairing — see wuji_writing_dataset.py:_assign_one_to_one_references for the canonical loader).

eval_refs/

Held-out ego videos for evaluation. Same LeRobot v2.1 schema as ego_ref/ (single observation.images.head camera, 720×1280, zero-filled 54-D state/action). 102 episodes total across 12 task_index values:

task_index	sub-task	top-level	ep	source
0	ball_left	ball	10	in-domain leftover (after 30 train picks)
1	ball_right	ball	10	in-domain leftover
2	cube_left	cube	10	in-domain leftover
3	cube_right	cube	10	in-domain leftover
4	cup_left_body	cup	5	in-domain leftover
5	cup_left_handle	cup	6	in-domain leftover
6	cup_right_body	cup	5	in-domain leftover
7	cup_right_handle	cup	5	in-domain leftover
8	bottle_left	bottle	10	OOD (not in training)
9	bottle_right	bottle	10	OOD
10	doll_left	doll	11	OOD
11	doll_right	doll	10	OOD
total			102

Per top-level category: ball/cube/bottle = 20, cup/doll = 21 (all ≥ 20).

Task_index 0..7 mirror the training tasks (so the same instructions in teleop/tasks.jsonl apply). Task_index 8..11 are new instructions for the OOD objects, listed in eval_refs/meta/tasks.jsonl.

This subset is intended as inference-time reference videos for VAM-style generation: feed an eval_refs clip + the matching task prompt to a model trained on teleop/+ego_ref/, then compare the generated robot video against held-out behavior (or qualitatively for OOD).

Loading

from lerobot.common.datasets.lerobot_dataset import LeRobotDataset

teleop = LeRobotDataset("yeeeiii111/wuji-pick-and-place", root="./data", subset="teleop")
ego    = LeRobotDataset("yeeeiii111/wuji-pick-and-place", root="./data", subset="ego_ref")
evalrf = LeRobotDataset("yeeeiii111/wuji-pick-and-place", root="./data", subset="eval_refs")

Or read meta directly:

import json
info = json.load(open("teleop/meta/info.json"))
episodes = [json.loads(l) for l in open("teleop/meta/episodes.jsonl")]

Provenance

file	meaning
<subset>/meta/info.json	LeRobot info (features, fps, totals, codec)
<subset>/meta/episodes.jsonl	per-episode entry (episode_index, tasks, length)
<subset>/meta/tasks.jsonl	the 8 task strings
<subset>/meta/episodes_stats.jsonl	per-episode min/max/mean/std for state/action and per-channel video stats
<subset>/data/chunk-000/episode_NNNNNN.parquet	per-frame state, action, timestamp, indices
<subset>/videos/chunk-000/<cam_key>/episode_NNNNNN.mp4	per-camera video

Limitations

• cup_left_handle has only 14 episodes (one fewer than the other 7 sub-tasks at 15) due to data collection.
• ego_ref state/action columns are zero-filled placeholders — only the head video carries information.
• teleop and ego_ref episodes are NOT temporally aligned (different recording sessions); they share task_index only.
• Sub-task labels for teleop were partially manual: left/right hand auto-detected from arm joint motion, but cup body vs handle was manually labeled by inspecting head-camera thumbnails.
• eval_refs videos were re-encoded from raw GoPro 4K to 720×1280 h264 (same as ego_ref).