arturaah/ur5_place_and_pour_nuts_camera_shifts_depth_raw_sidecar

UR5 Place-and-Pour Nuts — raw uint16 depth sidecar

Companion sidecar to arturaah/ur5_place_and_pour_nuts_camera_shifts. This repo contains raw RealSense D435i uint16 (millimetre) depth for every frame of every episode, in a layout that aligns 1:1 with the partner LeRobot v2.1 dataset's (episode_index, frame_index) indexing.

Why this exists

The observation.images.*_depth features in the partner dataset are lossy: they were normalized from raw Z16 to uint8 (clipped to 0–5 m), triplicated across RGB channels, then re-encoded as AV1 yuv420p video. Metric depth cannot be recovered from those videos.

This sidecar is the raw uint16 mm Z16 sensor depth re-extracted from the original ROS 2 rosbag recordings, intended for:

• GT-supervised distillation heads (e.g. π³ / Pi3X point/depth heads)
• Mix-supervised setups blending sensor depth with model-predicted depth
• Per-pixel point-cloud reconstruction when combined with the per-frame camera intrinsics K and extrinsics T_base_camera (in the partner dataset's pi3_scene_capture/ directory — calibration ships separately)

Contents

A single tarball pi3_scene_capture_depth_raw.tar that extracts to a directory tree mirroring LeRobot's video chunk layout:

pi3_scene_capture/
└── depth_raw/
    ├── chunk-000/
    │   ├── observation.images.context_left_depth/
    │   │   ├── episode_000000.npz
    │   │   ├── episode_000001.npz
    │   │   └── ...   (118 files)
    │   ├── observation.images.context_top_depth/
    │   └── observation.images.wrist_right_depth/
    └── manifest.yaml

354 .npz files total (118 episodes × 3 depth cameras). Each contains two arrays:

Key	Shape	Dtype	Meaning
depth	(T, 480, 640)	uint16	Raw RealSense Z16. Units = millimetres. 0 = invalid return (occluded / out-of-range / specular). Typical valid range 250–10000 mm; the sensor saturates around 10 m.
valid_mask	(T,)	bool	False on frames where no depth message fell within the 500 ms sync window of the master RGB frame; the corresponding depth[i] row is then all-zero. Mirrors the per-frame contract LeRobot uses.

T is identical to the partner LeRobot episode's length (asserted per-episode against meta/episodes.jsonl during extraction).

Camera-to-arm mapping (important!)

The friendly camera names are inverted vs the arm they're mounted on:

Feature key	Mount	Notes
observation.images.wrist_right_depth	LEFT arm (manipulator) wrist	Per-frame extrinsic = T_base_tool0 · T_tool0_cam (left-arm FK × hand-eye)
observation.images.context_left_depth	RIGHT arm (scene) wrist	Per-episode extrinsic — right arm is parked at one of 10 viewpoint poses. See scene_viewpoints.yaml + camera_positions_right.yaml.
observation.images.context_top_depth	world (fixed)	Single static T_base_camera

Extraction details

This session was a _camera_shifts study with 10 scene-arm viewpoints, 12 non-blacklisted episodes per viewpoint (the last 2 stayed at viewpoint 10). After 24 operator-flagged blacklist episodes and 4 conversion failures (corrupted zstd bags), 118 episodes survived end-to-end.

Sync: master RGB camera is context_top_rgb. Each depth frame is the nearest-in-time depth message within ±500 ms of the master frame's stamp (same rule as the LeRobot converter at src/record_scripts/rosbag_to_lerobotv21.py:665-691). Frames with no depth in window get valid_mask[i]=False and depth[i]=0.

Loading

import numpy as np

ep = np.load("pi3_scene_capture/depth_raw/chunk-000/"
             "observation.images.context_top_depth/episode_000000.npz")
depth_mm = ep["depth"]        # (T, 480, 640) uint16, units mm
valid    = ep["valid_mask"]   # (T,)         bool

# Match to the partner dataset's row index:
# lerobot[(episode_index, frame_index)] <-> depth_mm[frame_index]

License

Apache-2.0, same as the partner LeRobot dataset.