datasnack/brian_does_cleaning

DataSnack: Brian Does Cleaning (Yes, For Science)

Real-world household manipulation data captured by an actual human — Brian — using an iPhone 15 Pro with ARKit + LiDAR. No lab. No robot arm bolted to a table. Just a guy, a can of Windex, and a Nespresso machine, captured at 1920×1440 with lossless depth frames because your VLA deserves better than a blurry YouTube video.

Everything in this dataset is structured as LeRobot v3. Each session ships with a fully-formed LeRobot v3 dataset inside transformed.zip — observation.state, observation.images.rgb, observation.images.depth, action, subtask labels, language annotations. load_dataset() and go.

We're capturing 75+ sessions per week and pushing them here as they clear QA. Check back often — this dataset grows fast.

This dataset is part of the DataSnack capture pipeline: iPhone → transform → LeRobot v3. The goal is scene diversity at scale. Your model has seen one kitchen a thousand times. Now it can see a different one.

---

What's in here

3 sessions so far (more added weekly) — all from the same real home environment:

Session	Task	Duration	Frames
6C7B49D8	Windex (window/surface cleaning)	40.5s	~404
5C9E1CB4	Windex 2 (second take)	38.1s	~377
31FD1797	Nespresso (espresso machine operation)	142.3s	~1423

All sessions captured on an iPhone 15 Pro (LiDAR equipped), landscape-left orientation, in a domestic environment (San Mateo, CA).

---

LeRobot v3 format

Each session's transformed.zip contains a complete LeRobot v3 dataset ready to load:

from datasets import load_dataset
ds = load_dataset("datasnack/brian_does_cleaning")

LeRobot v3 features included per session:

Feature	Description
observation.images.rgb	1920×1440 HEVC @ 30fps
observation.images.depth	16-bit LiDAR depth @ 10fps
observation.images.dynamic_mask	Foreground mask aligned to depth frames
observation.state	6-DoF camera pose (tx, ty, tz, qx, qy, qz, qw)
observation.scene_state	ARKit scene mesh snapshot
action	Delta pose between frames
task	Natural-language task annotation (e.g. "clean surface with Windex")
subtask_labels	Per-episode subtask breakdown

The lerobot_v3/ directory inside each zip follows the standard LeRobot v3 structure with data/, meta/, and videos/ subdirectories.

---

File structure

Each session folder contains:

{session_id}/
├── rgb_video.mp4           # 1920×1440 HEVC @ 30fps, full session
├── hand_tracking.json      # Per-frame hand joint positions + confidence
├── scene_mesh.ply          # Static background mesh from ARKit
├── dynamic_mesh.json       # Dynamic foreground mesh (hands + objects)
├── foreground_cloud.json   # Point cloud of dynamic foreground content
└── transformed.zip         # LeRobot v3 dataset + lossless depth + metadata:
    ├── lerobot_v3/             # ← Full LeRobot v3 dataset (load_dataset() ready)
    │   ├── data/
    │   ├── meta/
    │   └── videos/
    ├── manifest.json           # Session summary (frame count, duration)
    ├── calibration.json        # Camera intrinsics (RGB + depth), extrinsic
    ├── conventions.json        # Coordinate system, depth units, formats
    ├── timestamps.csv          # Per-frame timestamps
    └── depth_frames/
        └── *.png               # Lossless 16-bit depth @ 10fps (256×192)

---

Sensor specs

Sensor	Spec
RGB	1920×1440, 30fps, HEVC
Depth (LiDAR)	256×192, 10fps, 16-bit PNG (millimeters)
Depth source	ARKit smoothedSceneDepth
IMU	100Hz (in raw capture; not in transformed output)
Coordinate system	ARKit right-handed, Y-up
Depth–RGB alignment	Registered; extrinsic in calibration.json

Depth values are lossless 16-bit unsigned integers in millimeters. Use depth_frames/*.png — the depth mp4 is 8-bit quantized and lossy.

---

Quality

Every session passed DataSnack's real-time QA layer during capture:

Metric	Windex	Windex 2	Nespresso
Tracking normal %	100%	100%	100%
Depth in optimal range %	100%	100%	100%
Mean depth high-conf ratio	75.2%	74.0%	93.9%
Hand visible %	87.4%	86.7%	96.1%
Mean ambient lux	989	995	1006

No post-hoc filtering. If it's in the dataset, it passed live QA at capture time.

---

Capture cadence

We're targeting 75+ sessions per week across household manipulation tasks — cleaning, kitchen prep, workshop tasks, and more. Sessions are pushed here as they clear the transform + QA pipeline, typically within hours of capture.

Star or watch this repo to get notified when new sessions drop. Or just check back — there will be more.

If you need a specific task type or environment, request it at datasnack.ai/contact.

---

What DataSnack is

DataSnack captures real-world manipulation data for physical AI training. The pipeline:

1. iPhone capture — ARKit + LiDAR, real-time QA, language-annotated episodes
2. Transform — depth registration, mesh reconstruction, hand tracking, LeRobot v3 packaging
3. Publish — LeRobot v3 native, with RLDS/TFRecord and HDF5 export also available

The thesis: scene diversity is the generalization bottleneck for VLAs. Labs produce high-rep, low-diversity data. DataSnack produces low-rep, high-diversity data — real homes, real workshops, real kitchens. Your model needs both.

---

License

CC BY 4.0 — use it, fine-tune on it, publish results. Credit DataSnack.

---

Citation

@dataset{datasnack_brian_does_cleaning_2026,
  author    = {DataSnack},
  title     = {DataSnack: Brian Does Cleaning (Yes, For Science)},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/datasnack/brian_does_cleaning}
}