Access to this dataset is reviewed manually. Please describe your intended use and affiliation. Requests are typically answered within TBD business days.

Pico Robotics Dataset · Annotated Edition

Egocentric multimodal capture from a Pico VR headset + custom tracker rig — Annotated tier

Builds on the Advanced edition by adding coarse action segmentation. Every sequence is divided into labelled temporal segments, so the data can be used directly for action recognition, temporal segmentation, and behaviour-understanding tasks without an annotation pass of your own.

🔒 This is a gated dataset. Access requests are reviewed manually; submit one from the dataset page.

Editions

Edition	Contents	Access
Basic	Undistorted stereo video + per-frame depth maps + dual-channel audio + ~1 kHz 6-DoF head pose + camera calibration	🟢 Public
Advanced	Everything in Basic + pose-aligned stereo point clouds (`.npz`) + 21-joint hand tracking + world-frame point clouds + MCAP / Foxglove visualization	🔒 Request access
Annotated (this repo)	Everything in Advanced + coarse action segmentation (`segments.json`)	🔒 Manual review

Scenes

Config	Environment	Description	Sequences	Segments	Duration
`retail-shelf-a`	Convenience store, storefront area	Beverage gondolas, snack shelving, chest freezer, liquor display; strong daylight through a glass facade with mixed indoor lighting	TBD	TBD	TBD
`retail-shelf-b`	Convenience store, interior aisles	Branded beverage coolers, free-standing snack racks, bottled-water pallets, tiled floor under uniform ceiling lighting	TBD	TBD	TBD
`gymnasium`	Indoor sports arena	Large open court with line markings, tiered seating, overhead truss lighting, banners, and event-setup activity	TBD	TBD	TBD

from datasets import load_dataset

ds = load_dataset("skycn110/pico-robotics-annotated", "retail-shelf-a")

Directory structure

retail-shelf-a/
└── sample_0001/
    ├── head_left_camera_undistorted.mp4
    ├── head_right_camera_undistorted.mp4
    ├── depth/
    ├── audio_dual_channel.wav
    ├── undistort_camera.json
    ├── video_index.json
    ├── pointcloud/
    ├── pointcloud_world/
    ├── hand_landmarks.json
    ├── recording.mcap
    └── segments.json                       # Coarse action segmentation  ← new in this tier

Everything except segments.json is documented in the Advanced README.

Annotation format

`segments.json`

{
  "sample_id": "sample_0001",
  "scene": "retail-shelf-a",
  "fps": 30,
  "segments": [
    {
      "id": 0,
      "start_frame": 0,
      "end_frame": 148,
      "start_time": 0.0,
      "end_time": 4.93,
      "label": "approach_shelf",
      "notes": ""
    }
  ]
}

Field	Type	Description
`start_frame` / `end_frame`	int	Inclusive frame range, indexed against `video_index.json`
`start_time` / `end_time`	float	Seconds from sequence start
`label`	string	Action class, see the label set below
`notes`	string	Free-text annotator remark, may be empty

Segments within a sequence are contiguous and non-overlapping; frames that fit no class are labelled other.

Label set

TBD — list the action classes here, with a one-line definition and the frame count for each, for example:

Label	Definition	Segments
`approach_shelf`	Operator walks toward a shelf until stationary in front of it	TBD
`scan_shelf`	Head sweeps across shelf contents without hand motion	TBD
`reach_and_grasp`	Hand extends toward a product and closes on it	TBD
`inspect_item`	Held item brought toward the camera and rotated	TBD
`place_back`	Held item returned to the shelf	TBD
`walk_transit`	Locomotion between areas	TBD
`other`	Anything not covered above	TBD

Annotation protocol


Annotators	TBD
Guidelines	TBD (link to the written protocol)
Boundary tolerance	TBD frames
Double-annotated portion	TBD %
Inter-annotator agreement	TBD

Labels are deliberately coarse: boundaries are approximate and the class vocabulary is small. They are intended as weak supervision or as a starting point for finer annotation, not as a precision benchmark.

Quick start

import json

with open("retail-shelf-a/sample_0001/segments.json") as f:
    seg = json.load(f)

for s in seg["segments"]:
    print(f"{s['start_time']:6.2f}–{s['end_time']:6.2f}s  {s['label']}")

Clip a segment with the video:

import cv2

cap = cv2.VideoCapture("retail-shelf-a/sample_0001/head_left_camera_undistorted.mp4")
s = seg["segments"][0]
cap.set(cv2.CAP_PROP_POS_FRAMES, s["start_frame"])

for _ in range(s["end_frame"] - s["start_frame"] + 1):
    ok, frame = cap.read()
    if not ok:
        break

Intended uses

Egocentric action recognition and temporal action segmentation
Weakly supervised pretraining for behaviour understanding
Video–language grounding of first-person activity
Segment-conditioned imitation learning

Out of scope

Class balance is uneven and follows whatever occurred naturally during capture; some labels have very few segments. Boundaries are coarse and were not adjudicated frame-by-frame. Reporting state-of-the-art numbers on this label set without acknowledging those limits would be misleading.

License

Released under CC BY 4.0. When using this dataset, please attribute:

Pico Robotics Dataset by skycn110, licensed under CC BY 4.0

Citation

@misc{skycn110_pico_robotics_annotated,
  title  = {Pico Robotics Dataset: Annotated Edition},
  author = {skycn110},
  year   = {2026},
  url    = {https://huggingface.co/datasets/skycn110/pico-robotics-annotated}
}

Contact

Data questions / collaboration: skycn110@gmail.com
Access requests: use the form on this dataset page (manually reviewed)
Issues and Discussions: welcome on this repository

Downloads last month: 12

Total file size:

9.53 kB