Staer Warehouses — v0.5

Synthetic stereo walkthroughs through procedurally-generated and curated warehouses, rendered with NVIDIA Isaac Sim. Each scene includes RGB, depth, semantic segmentation, instance segmentation, IMU, and a scene-graph JSON with per-asset class labels and oriented bounding boxes (OBBs) in world coordinates.

👉 Interactive preview: https://huggingface.co/spaces/staerrobotics/warehouses-viewer

What's new in this release

• Full per-pixel ground truth now ships with every walkthrough — 16-bit inverse depth, semantic segmentation, and instance segmentation as lossless FFV1 videos (depth0/, semantic0/, instance0/), plus IMU (imu0/data.csv). Previous releases included only the RGB stream and a lossy preview grid; the depth shown there was a visualisation, not data.
• Stereo — both cameras of the ZED-X-like rig (cam0/cam1) are now included per walkthrough.
• Structured scene descriptions — every scene ships with a scene_graph.json that enumerates each object in the warehouse with its class label, oriented bounding box (OBB) in world coordinates, axis-aligned bounding box, and pose (position + quaternion). One file per scene, shared across all walkthroughs of that scene.
• All authored trajectories shipped per scene — all 147 authored trajectories across 15 scenes, giving substantially more pose / viewpoint diversity for the same set of warehouse environments.

Contents

• 15 scenes, 147 walkthrough trajectories total
• Per walkthrough: stereo cam0/cam1, ~300–2400 frames at 20 fps (≈15s–2min camera paths)
• 1920×1080 H.265 RGB per camera; 16-bit inverse depth, semantic and instance ID maps as lossless FFV1 (gray16le) videos
• Per-scene scene_graph.json with OBB + class for every object
• IMU + ground-truth poses per camera, EuRoC-ASL convention

Scenes

#	Name	Walkthroughs	Objects	Top classes
01	scene01	9	33254	box(14128), crate(12614), pallet(2728)
02	scene02	11	11599	box(3742), crate(2421), pallet(1512)
03	scene03	10	14661	rack(4150), box(2430), pallet(2010)
04	scene04	9	18339	box(5067), rack(3240), crate(2713)
05	scene05	9	15653	box(4583), rack(3300), crate(2595)
06	scene06	12	4303	box(587), ceiling_light(448), rack(440)
07	scene07	8	15294	box(5341), crate(2850), pallet(2305)
08	scene08	8	5085	box(1655), crate(1041), pallet(534)
09	scene09	10	360	drum(150), jerrican(52), railing(36)
10	scene10	10	945	conveyor(417), belt_item(144), box(142)
11	scene11	10	471	box(206), pallet(66), crate(36)
12	scene12	11	441	box(177), belt_item(43), tote(36)
13	scene13	10	677	box(189), shelving(153), bin(107)
14	scene14	10	643	box(246), pallet(60), bin(54)
15	scene15	10	567	box(145), tote(66), drum(63)

Class taxonomy

51 unique classes across the release. Top 20 by count:

Class	Count
box	38,638
crate	26,046
pallet	14,809
rack	13,769
ceiling_light	4,620
pipe	3,588
wall	2,964
fire_equipment	2,312
wiring	2,236
ceiling	1,576
truss	1,320
floor	1,320
decor	1,172
camera	692
ventilation	660
phone	660
column	660
wall_light	576
electrical_panel	576
alarm	552

File layout

warehouse-v0.5/
├── README.md
├── overview.png              # 5×3 montage of scene thumbnails
├── scene01/
│   ├── thumbnail.jpg         # representative frame for the scene
│   ├── scene_graph.json      # objects[]: class, OBB, pose (per-scene)
│   └── walkthroughs/
│       ├── 001/
│       │   ├── preview.mp4   # 2×2 grid: RGB | depth | semantic | instance
│       │   ├── manifest.json # EuRoC-ASL header (timestamps, fps, frameCount)
│       │   ├── metadata.json # asset metadata (modalities, resolution, …)
│       │   ├── cam0/
│       │   │   ├── poses.csv   # T_WC (Hamilton w,x,y,z) at frame timestamps
│       │   │   ├── sensor.yaml # intrinsics, T_BS
│       │   │   ├── data.csv    # frame index: timestamp per video frame
│       │   │   └── video.mp4   # RGB (H.265)
│       │   ├── cam1/  …         # stereo right camera, same files
│       │   ├── depth0/
│       │   │   ├── video.mkv         # 16-bit inverse depth (FFV1 gray16le)
│       │   │   ├── depth_ranges.json # per-frame inv_min/inv_max (see below)
│       │   │   └── data.csv
│       │   ├── semantic0/
│       │   │   ├── video.mkv         # 16-bit class IDs (FFV1 gray16le)
│       │   │   ├── id_to_labels.json # class ID → label name
│       │   │   └── data.csv
│       │   ├── instance0/
│       │   │   ├── video.mkv         # 16-bit instance IDs (FFV1 gray16le)
│       │   │   ├── id_to_labels.json # instance ID → prim path
│       │   │   └── data.csv
│       │   └── imu0/
│       │       └── data.csv    # gyro + accel at 400 Hz, EuRoC columns
│       ├── 002/  …
│       └── NNN/  …
└── scene02/ … scene15/

scene_graph.json is shared across all walkthroughs of a scene (same underlying USD); cam poses differ per walkthrough. Depth and segmentation are rendered from cam0 (the left camera). Row i of each data.csv gives the timestamp of video frame i.

Decoding the 16-bit streams

video.mkv files are lossless FFV1, one 16-bit grayscale plane per frame — every pixel value is exactly the value the renderer wrote. Extract frames with ffmpeg or read them directly with PyAV:

ffmpeg -i depth0/video.mkv -f image2 -pix_fmt gray16be depth_%06d.png

import av, numpy as np
frames = [f.to_ndarray() for f in av.open('depth0/video.mkv').decode(video=0)]

Depth is stored as per-frame-normalised inverse depth. To recover metric depth for frame i with 16-bit value v (0 = invalid / no hit):

r = depth_ranges[i]  # depth_ranges.json, ordered like the video frames
inv = r['inv_min'] + (v - 1) / 65534.0 * (r['inv_max'] - r['inv_min'])
depth_m = 1.0 / inv  # metres

Semantic / instance pixel values are integer IDs; map them to labels via the adjacent id_to_labels.json.

Coordinate conventions

• Right-handed world frame, +Z up.
• Camera convention: T_WC = T_WB · T_BS where T_BS comes from sensor.yaml.
• Quaternions are Hamilton, (w, x, y, z), with positive-w convention.
• Distances in metres; timestamps in nanoseconds (EuRoC-ASL).

Rolling releases

This dataset is on a rolling-release cadence: minor versions ship regularly with quality improvements and bug fixes layered onto the previous release rather than waiting for a single large drop. The scene set and modalities are stable across minor versions; what changes is per-scene asset quality, trajectory authoring, and annotation coverage.

Known issues we are actively addressing in subsequent minor releases:

• Trajectory coverage — a small number of authored walkthroughs are still missing from individual scenes and will be backfilled in the next minor release.
• Asset material quality — a subset of NVIDIA SimReady props ship with incomplete MDL/texture chains that fall back to default (white or near-black) shaders; we are auditing and patching these per scene.
• Lighting consistency — the curated NVIDIA themed shells (scenes 09–15) have noticeably dimmer baked lighting than the procedural scenes (01–08); we are normalising exposure across the set.
• Trajectory smoothness — a few canvas-authored trajectories include sharp orientation changes that can confuse motion estimators; we are re-authoring with stricter smoothness constraints.

Subscribe to the dataset on Hugging Face to be notified when a new minor version is published. Each release ships with its own version tag and a short changelog summarising the deltas from the previous release.

Citation

If you use this dataset, please cite Staer Robotics.

@dataset{staer_warehouses_v0.5, title={Staer Warehouses (v0.5)}, year={2026}, publisher={Staer Robotics}, url={https://huggingface.co/datasets/staerrobotics/warehouses}}}