simpk/single_pickplace

pick_and_place-300 — Unitree G1 + Dex3, "pick octopus and place inside brown basket"

LeRobot v2.1 dataset. Teleoperated bimanual G1 with Dex3 hands; lower body under a GR00T whole-body-control policy, upper body teleoperated.

Episodes	349 (322 positive demos + 27 negative samples)
Frames	161,440 (2.24 h @ 20 fps)
FPS	20
Cameras	3 × h264 640×480 yuv420p
State / action	43-dim whole body (float64)
Task string	pick octopus and place inside brown basket
Size	2.5 GB

Cameras

Key	Mount	Maps to (pi0.5)
observation.images.ego_view	head/chest, looks down at the table; sees both hands, object and basket	base_0_rgb
observation.images.ego_left	left wrist	left_wrist_0_rgb
observation.images.ego_right	right wrist	right_wrist_0_rgb

Joint layout — read this before training

observation.state and action are 43-dim whole-body vectors:

Indices	Group	Notes
0–11	legs	WBC-controlled, not teleoperated
12–14	waist	enable_waist: false during collection
15–42	upper body — the 28 dims to train on	contiguous, so a slice is enough

The 28 upper-body dims are ordered [L_arm 7, L_hand 7, R_arm 7, R_hand 7] — not [arm 14, hand 14]. Any transform that groups arms against hands must interleave: a delta-action mask is make_bool_mask(7, -7, 7, -7), and make_bool_mask(14, -14) would leave the right arm (the one performing the task) on absolute actions.

Data properties that look like bugs but are not

• 27 episodes are deliberate negative samples: 51–63, 65–74, 76–79. The octopus, the basket, or both are absent from the table, so the robot idles and the hand is never commanded. Right-arm motion is 0.269 rad peak-to-peak against 1.649 for real demos. Keep them — they teach the policy not to act on an incomplete scene.
• left_hand_* action is bit-identically 0.0 across all 161,440 frames (all 7 dims). The left hand was never commanded; 0.0 is open/neutral. 7 of the 28 action dims are constant.
• Coupled fingers: right_hand_index_1 == right_hand_middle_1 and right_hand_thumb_1 == right_hand_thumb_2, bit-identical. Real right-hand DoF is 5, not 7.
• Hand actions overshoot the reachable range — right_hand_middle_0 is commanded to 2.98 rad against a reachable 1.56 (smooth_hand_grasp: true). The overshoot is the grip force; clamping to the observed state range would silently weaken every grasp. Clamp to the training action envelope instead (scripts/clamp_spec.json).
• 7 recording sessions, boundaries at episode indices 0, 80, 131, 174, 184, 311, 338. They differ in table position, basket (session 4 uses a lighter wicker one), lighting and left-arm rest pose (session 0 notably). Subsample across sessions, not by prefix.

Known defects

• Episodes 172, 173 — full, successful executions with all three cameras frozen on a single still frame. 698 frames that teach manipulation succeeding with zero visual change.
• Episodes 146, 147 — ego_left frozen for 240 / 402 frames; the other two cameras are fine.

These are retained for completeness. observation.img_state_delta is a reliable detector: normal is ~0.033 s, these read 21–595 s. Exclude them with scripts/make_subset.py if you want a strictly clean visual set.

Missing mode

There is no post-completion hold. Positive demos show a median of 5 frames (0.25 s) of trailing stillness; only 1% reach 1 s and none reach 2 s. After release the arm retreats. A policy trained on this alone has no signal to stop after success and will tend to re-attempt the task.

Provenance note

This dataset is 7 recording sessions concatenated. The files were renamed contiguously but the in-parquet episode_index / index columns were originally left at their per-session values, which made 269 of 349 episodes report the wrong episode_index. Because LeRobot uses that column to pick the video file and to clamp action-chunk boundaries, 77% of samples would have drawn images from the wrong episode — silently, with a healthy-looking loss curve. This is repaired in the published version (scripts/repair_indices.py, idempotent, --dry-run supported). meta/episodes.jsonl, meta/episodes_stats.jsonl and all 1047 videos were already aligned to filenames and are unchanged.

Contents

data/chunk-000/episode_*.parquet     349 files
videos/chunk-000/observation.images.{ego_view,ego_left,ego_right}/episode_*.mp4
meta/{info,modality}.json  meta/{episodes,episodes_stats,tasks}.jsonl
PI05_G1_PICKPLACE_RUNBOOK.md         end-to-end pi0.5 SFT runbook -- START HERE
PI05_G1_DEX3_SFT_RUNBOOK.md          original runbook (superseded; written for a different dataset)
RUNBOOK_CORRECTIONS.md               what changed between the two, and why
g1_dex3_policy.py                    openpi transforms (43->28 slice, camera mapping)
openpi_config_block.py               openpi TrainConfig + DataConfig to paste in
scripts/repair_indices.py            index/counter repair (already applied)
scripts/make_subset.py               build stratified episode subsets for scaling studies
scripts/test_transforms.py           verifies the transform pipeline against real rows
scripts/check_token_len.py           verifies max_token_len is large enough for pi0.5
scripts/verify_download.py           checks a downloaded copy is complete and consistent
scripts/patch_openpi_config.py       applies the policy + train configs to an openpi checkout
scripts/check_dataloader.py          pulls one sample through the real pipeline before training
scripts/prune_checkpoints.py         strips train_state from old checkpoints (600 GiB -> 188 GiB)
scripts/clamp_spec.json              per-joint robot-side clamp limits
scripts/upload_to_hf.py              publish to the Hub (tags v2.1, guards on index integrity)
reference_frames/                    frame 0 per recording session, for scene setup

filter_upper.py, add_quantiles.py and add_image_stats.py at the root target a different pipeline (LeRobot v3.0 + lerobot-native pi0/pi05 training). They are not used by the runbook here and will not run against a v2.1 layout. Kept for reference.

Privacy

Recorded in a shared office. Bystanders are visible and identifiable in the wrist-camera views.