Request access to the Suralto construction sample
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
This sample is shared for evaluation under the Suralto sample licence (see LICENSE.md). Tell us a little about your work and we will review your request.
Log in or Sign Up to review the conditions and access this dataset content.
Suralto Egocentric Construction: Foundation Work
Real skilled-trade work, seen through the worker's eyes, structured for Physical AI.
This sample is 10.7 minutes of first-person video from an active construction site: a rebar worker tying cages, placing bars and bending dowels during normal foundation work. Every second is segmented into task steps, the gloved hands are segmented frame by frame, and every frame carries depth. The footage is consented, de-identified and delivered clean, with all annotations as separate, time-aligned files.
| Footage | 10.7 min continuous, head-mounted, first-person |
| Trade | Foundation: rebar cage tying, bar placement, dowel bending |
| Video | 1080p H.264, 30 fps, no audio |
| Source resolution | 3840x2160 (available on request) |
| Task steps | 65 time-stamped steps covering 100% of the video |
| Hands | pixel masks + left/right on gloved hands, found in 86% of 19,170 frames |
| Depth | per-frame relative depth + 32 point clouds |
| Privacy | faces and plates blurred and human-checked, audio removed, GPS stripped |
| Quality | 14 automated QA checks passed; SHA-256 for every file |
| Metadata | Croissant 1.0 with Responsible-AI fields, full provenance |
Why this data
- Real work, not staged. Recorded during normal work on a live site, with the clutter, gloves, tools and interruptions that robots will actually face.
- Task structure, not just pixels. Each step has a start and end time, the tool in hand, which hand does the work, a verb from a fixed action list, the object, the outcome and any safety flag.
- Gloved hands, tracked. Off-the-shelf hand trackers do not see gloved work hands: on this footage MediaPipe Hands found none. Our pipeline segments gloved hands in 86% of frames, with a mask outline, box and consistent left/right identity per frame.
- 3D context. Relative depth for every frame and point clouds every 20 seconds give scene geometry for world models and manipulation policies.
- Model-ready video. Nothing is drawn on the footage. Hands, depth and labels ship as separate files you can join on time.
- Consent and privacy by design. Signed contributor consent, faces and licence plates blurred and then checked by a person, audio removed, GPS and device metadata stripped, people and sites identified only by ID.
- Auditable.
provenance.jsonrecords the source checksum and every processing step;qa_report.jsonlists every automated check; every file has a SHA-256.
Files
| Path | Contents |
|---|---|
video/segment_1080p.mp4 |
the footage, 1080p, no audio, no overlays |
labels/steps.csv, labels/steps.json |
65 task steps (also browsable in the viewer above) |
3d/hands.jsonl, 3d/hands_meta.json |
per frame: each hand's mask polygon, box, score and left/right |
3d/depth.mp4, 3d/depth_small.npz |
relative monocular depth, every frame |
3d/cloud_*.ply, 3d/clouds.json |
point clouds at keyframes (every 20 s) |
provenance.json, qa_report.json |
consent ID, source checksum, processing history, QA results |
croissant.json, SHA256SUMS |
Croissant 1.0 + RAI metadata, file checksums |
viewer.html |
open in Chrome and choose this folder to watch video, steps, hands and depth together |
Step schema
| Field | Values |
|---|---|
start_s, end_s |
seconds from the start of the video |
step |
short description in trade language |
tool |
tool in hand, or none |
hands |
both, left, none, right |
action |
carry, cut, drill, grasp, hammer, hold, inspect, level, lift, mark, measure, other, place, pour, spread, tie, twist, walk |
object |
what is acted on |
outcome |
correction, error, success, unclear |
safety |
near_miss, none, other, ppe_missing, unsafe_posture, unsafe_tool_use |
First steps:
| start s | end s | step | tool | hands | action |
|---|---|---|---|---|---|
| 0.0 | 8.1 | twist and cut tie wire on horizontal rebar with pliers | rebar pliers | both | twist |
| 8.1 | 21.2 | step onto circular rebar cage and adjust tie wire position | rebar pliers | both | place |
| 21.2 | 27.8 | step across cage and prepare pliers for tying | rebar pliers | both | walk |
| 27.8 | 39.4 | twist tie wire at rebar intersection on cage | rebar pliers | both | twist |
| 39.4 | 44.2 | reposition feet and pliers along cage frame | rebar pliers | both | walk |
| 44.2 | 54.1 | twist and tighten tie wire on cage ring | rebar pliers | both | twist |
Quick start
import json
import pandas as pd
steps = pd.read_csv("labels/steps.csv") # one row per task step
hands = [json.loads(line) for line in open("3d/hands.jsonl")] # one row per video frame
frame = hands[300] # hands at frame 300
print(steps[["start_s", "end_s", "step", "action"]].head())
print(frame["t"], [(h["hand"], h["box"]) for h in frame["hands"]])
Intended uses
Robot manipulation and imitation learning, vision-language-action and world-model pretraining, egocentric activity recognition and step segmentation, hand-object interaction, and construction safety research.
How it was made
- Recorded by the worker on a head-mounted camera during normal work, with signed consent.
- Trimmed; device metadata and GPS removed.
- Faces and licence plates blurred at full resolution (EgoBlur Gen2), then checked by a person; audio removed.
- Step labels generated by Gemini on Vertex AI in fixed vocabularies.
- Gloved hands detected with OWLv2 and segmented with SAM 2.1; depth from Depth Anything V2 Small. All models used are licensed for commercial use.
- Automated QA and packaging with provenance and Croissant metadata.
Limitations
Step labels are machine-generated (Gemini on Vertex AI) in a fixed vocabulary. The sample covers one site, one trade and one camera. Depth is relative, not metric. Hand masks can miss hands that are heavily motion-blurred or hidden behind tools.
Licence and access
Shared for evaluation under the Suralto sample licence (LICENSE.md): you may inspect it and evaluate models
internally. Redistribution, attempts to identify any person, and commercial training require a written
agreement with Suralto. Request access with the form on this page.
About Suralto
Suralto builds consented, expert-structured egocentric datasets of skilled trades for Physical AI. This sample is one clip from our pipeline; larger volumes, other trades, native-resolution video and human-reviewed labels are available on request. Mention your needs in the access request and we will follow up.
- Downloads last month
- 99