benudavis/cluttered_grasp

Cluttered Grasp (UR10 + Robotiq 2F-140 + tactile + depth)

Real-robot teleop demonstrations of cluttered grasping. Episodes are gated on gripper_tcp height in base_link (record while low, pause while high), time-synced to the RealSense RGB master @ 30 Hz, then converted to LeRobot format.

Robot	UR10 + Robotiq 2F-140 (ur10_robotiq_2f140)
FPS	30 (RGB master timeline)
Episodes / frames	See meta/info.json (total_episodes, total_frames)
Codebase	LeRobot dataset v3.0
Task	cluttered grasping
Recovery subset	Last 284 episodes (episode_index 710–993): 90 + 194 dedicated recovery trajectories. The rest are mixed.

Sensors & resolutions

Stream	Capture (ROS)	Dataset feature	Shape (H×W×C)	Notes
Center RGB	RealSense color 1280×720@30 JPEG (color/image_raw)	observation.images.center_cam	720×1280×3	16:9 color optical frame
Depth	RealSense depth module 848×480@30 (W×H), /camera/camera/depth/image_rect_raw — not align-to-color	observation.images.center_cam_depth	480×848×1 (uint16 mm, video12)	Unaligned native depth grid; not registered to RGB pixels
Tactile L	Fingertip cam JPEG (~90 Hz capture, hold-last to RGB)	observation.images.tactile_L	240×320×3	4:3
Tactile R	same	observation.images.tactile_R	240×320×3	4:3

• Master clock: compressed RGB frames — one dataset row per RGB stamp (~30 Hz).
• Sync: nearest-neighbor within ~50 ms, then hold-last for depth/tactile/proprio.
• Depth (logged & exported): native D4xx depth at 848×480 from depth/image_rect_raw into bags → zarr center_cam_depth → LeRobot. Unit millimetres (uint16). Not the 1280×720 aligned_depth_to_color stream (that is used at runtime for ContactGraspNet only).
• Depth vs RGB: depth rows are time-synced to RGB but not pixel-aligned — fuse with extrinsics / reprojection, not naive (u,v) overlay on center_cam.
• Depth export for this dataset: 12-bit gray12le HEVC (video12), log-quantized over 0.2–1.0 m.

Dataset structure

cluttered_grasping/
├── meta/
│   ├── info.json          # schema, fps, totals, video codec info
│   ├── stats.json         # per-feature normalization stats
│   ├── tasks.parquet      # task prompts
│   ├── episodes/          # per-episode metadata
│   └── tactile_refs/      # pre-trim tactile Δ baselines (one .npz per episode)
│       ├── 000000.npz
│       ├── 000001.npz
│       └── …              # {episode_index:06d}.npz
├── data/
│   └── chunk-*/file-*.parquet   # state, action, indices
└── videos/
    ├── observation.images.center_cam/chunk-*/file-*.mp4        # AV1 1280×720 (W×H)
    ├── observation.images.center_cam_depth/chunk-*/file-*.mp4  # HEVC gray12le unaligned depth 848×480 (W×H)
    ├── observation.images.tactile_L/chunk-*/file-*.mp4         # AV1, 320×240
    └── observation.images.tactile_R/chunk-*/file-*.mp4

Tactile reference sidecars (meta/tactile_refs/)

Training subtracts a per-episode tactile baseline before letterboxing (Δ tactile). LeRobot video rows are leading-trimmed at export (home hold / NaN policy dropped), but the ref is computed from the raw source zarr at frame 0 so it matches labeling, deploy, and live recording (mean of the first 10 non-empty tactile frames at bag start).

One compressed NumPy archive per LeRobot episode_index:

Path	Contents
meta/tactile_refs/{episode_index:06d}.npz	Per-episode tactile Δ reference

Each .npz file:

Key	dtype	shape	Description
l	uint8	H×W×3	Mean left tactile RGB ref (native ~240×320)
r	uint8	H×W×3	Mean right tactile RGB ref (native ~240×320)
leading_trim_drop	int32	scalar	Frames dropped from zarr start at LeRobot export (NaN + stationary EE)

Not a LeRobot parquet/video feature — load sidecars by episode_index when preprocessing tactile (e.g. cluttered-grasp-planner training/tactile_refs.py). Written on full convert and via zarr_to_lerobot.py --refs-only (no video re-encode). Include meta/tactile_refs/ on HF upload if you train with sidecar-based Δ tactile.

Features (per frame)

Key	Type	Shape	Description
observation.state	float32	(8,)	eef_x/y/z, eef_qx/qy/qz/qw (xyzw), measured gripper at observation t
action	float32	(7,)	delta_x/y/z, delta_roll/pitch/yaw (world, t→t+1) + gripper mode {-1,0,+1} at t (last frame dropped)
observation.joint_state	float32	(7,)	joint_0..5, measured gripper at t (gripper redundant with observation.state)
observation.joint_action	float32	(7,)	joint_0..5 absolute at t+1 + gripper mode at t (same mode as action)
observation.gripper_setpoint	float32	(1,)	Absolute finger_joint command at t (under-the-hood; not the discrete action)
observation.cgn_grasp_pose	float32	(7,)	Constant CGN TCP target in base_link (xyz + quat xyzw); broadcast every frame
episode_state	int64	(1,)	Trajectory phase: 0 reach (home→CGN pose), 1 grasp_lift
intervention	int64	(1,)	Sticky 1 after human teleop starts this episode (else 0)
grasp_state	int64	(1,)	Exclusive grasp phase at observation t: 0 no_contact, 1 contact, 2 unstable_closure, 3 stable_closure, 4 lift, 5 success
grasp_sequence	int64	(1,)	Episode-level closure-chain class, same id on every frame. 0 other, 1 scl, 2 ucl, 3 ucl→scl, 4 ucl→ucl, 5 scl→ucl, 6 three_closure. See package README.
next.success	float32	(1,)	1.0 on the last exported frame when terminal grasp_state is success; else 0
next.done	float32	(1,)	1.0 on the last exported frame of every episode; else 0
observation.images.center_cam	video (AV1)	720×1280×3	Scene RealSense RGB (1280×720 capture)
observation.images.center_cam_depth	video (HEVC gray12le)	480×848×1	Unaligned native depth 848×480 capture (H×W×C storage); mm; is_depth_map
observation.images.tactile_L	video (AV1)	240×320×3	Left fingertip tactile (/tactile1)
observation.images.tactile_R	video (AV1)	240×320×3	Right fingertip tactile (/tactile2)
timestamp	float32	(1,)	Time within episode (s)
frame_index / episode_index / index / task_index	int64	(1,)	LeRobot indices

observation.state EE pose is gripper_tcp in base_link.
action pose delta: delta_xyz = pos[t+1]-pos[t]; delta_rpy = euler_xyz(R[t+1] @ R[t].T) (world/fixed-frame).
Gripper on both action channels is discrete mode (-1 open, 0 hold, +1 close), not absolute position. Under the hood teleop/CGN ramp a setpoint (logged in zarr as gripper_setpoint) with the same controller.
Wrist force/torque is not exported to LeRobot (still present in upstream zarr if needed).

Label example: terminal phases … success, success → next.done = 1 on the last exported frame; next.success = 1 on that frame too. Failed episodes still get next.done = 1 on the last frame; next.success = 0.

Episode definition

• Start: cgn_teleop_trial accepts a grasp preview and calls begin_episode (bag warm-up at home, then approach).
• End: EE reaches lift handoff (LIFT_Z for teleop, or after auto lift); end_episode then place/home are not recorded.
• No recording between trials.
• Fully automated vs human: check intervention max / zarr attr intervention_used.

Episode composition

Source zarrs are converted in sorted path order, so later collection days sit at the end of the LeRobot index range.

Episodes (episode_index)	Count	Collection	Content
0–709	710	2026-08-11 … 2026-08-25	Mix of everything (auto + teleop, success + fail, etc.)
710–799	90	2026-08-28	Recovery trajectories (explicit recovery collection)
800–993	194	2026-09-01	Recovery trajectories (explicit recovery collection)

The last 90 + 194 = 284 episodes are recovery-only. Slice with episode_index >= 710 (or 710:800 / 800:994 for the two recovery days). Earlier episodes are not recovery-tagged; they mix all trial types.

Load

from lerobot.datasets.lerobot_dataset import LeRobotDataset

ds = LeRobotDataset("benudavis/cluttered_grasp")
print(ds)
frame = ds[0]
# EE state / action:
#   frame["observation.state"]         # (8,) eef xyz+quat + gripper
#   frame["action"]                   # (7,) delta_xyz + delta_rpy + abs gripper next
#   frame["observation.joint_state"]  # (7,) joints + gripper
#   frame["observation.joint_action"] # (7,) next joints + gripper
#   frame["observation.cgn_grasp_pose"]  # (7,) constant CGN target
# RGB / tactile: frame["observation.images.center_cam"], ...
# Depth: frame["observation.images.center_cam_depth"]  # 480×848×1 mm, unaligned 848×480 capture
# Phases / returns:
#   frame["episode_state"]    # 0 reach / 1 grasp_lift
#   frame["intervention"]     # 1 if teleop used
#   frame["grasp_state"]      # int64 phase id (contact labels)
#   frame["grasp_sequence"]   # episode class: 0 other, 1 scl, 2 ucl, 3 ucl→scl, …
#   frame["next.success"]     # 1.0 on last exported frame if episode ends in success
#   frame["next.done"]        # 1.0 on last exported frame (every episode)

Collection / conversion notes

• ROS bags → one zarr episode per bag (bag2dataset under ~/venvs/clutter, zarr≥3).
• Episodes are service-gated by cgn_teleop_trial (begin after preview Enter with warm-up; end at lift handoff).
• Offline rebuild: scripts/zarr_to_lerobot.py --fps 30 --depth-mode video12 over ~/EDG_Experiment/clutter/**/*.zarr (skips clutter/old/).
• Phase labels: scripts/label_grasp_states.py (Streamlit) writes data/obs/policy/grasp_state and, on Save / Save all, derived next_success / next_done on each episode zarr.
• Converter sets next.done / next.success with the same terminal-frame rules as the labeler (last frame; success from terminal grasp_state). grasp_sequence is derived from the trimmed closure chain (see package README). Existing datasets: zarr_to_lerobot.py --labels-only patches labels + adds grasp_sequence without re-encoding videos. zarr_to_lerobot.py --refs-only writes or refreshes meta/tactile_refs/*.npz only.
• Trial stats: python scripts/plot_trial_outcomes.py --zarr-root … --out … (or zarr_to_lerobot.py --trial-stats-out …).
• Bag convert also skips leading frames with empty tactile / NaN policy before writing zarr.
• Auto-label may set contact/success from tactile blob area; only terminal success runs (touching episode end) stay success.
• Actions: EE delta pose + absolute next gripper; joint absolute next (see scripts/lerobot_convert/actions.py).
• Upload: scripts/upload_lerobot_hf.py (copies this card to the dataset root as README.md).