Worlddatalabs/egocentric-manufacturing

Egocentric Manufacturing

High-diversity factory egocentric video for VLA and industrial robotics — first-person capture from live manufacturing floors, not staged demos, not recycled datasets.

Collected, curated, and licensed by World Data Labs.

Overview

Egocentric Manufacturing is a first-person video corpus captured on live manufacturing floors. Operators wear headset cameras while performing authentic plant work — assembly, material handling, packing, bagging, filling, palletizing, quality inspection, and warehousing. Every clip ships with complete scene segmentation: ordered subtasks with skill labels, per-step outcomes, an entity inventory, and narration aligned to the source timeline.

Frontier physical-AI models are bottlenecked on real egocentric manufacturing footage. Simulation and open web video miss factory lighting, hand–tool interaction, cluttered workcells, and long-horizon process structure. This corpus is built for:

• Video pretraining and continued pretraining on long-horizon industrial footage
• VLA and visuomotor policies — perception towers and temporal encoders for factory skills
• Skill and process understanding, supervised directly by the subtask annotations
• Domain adaptation from open egocentric corpora onto factory floors
• Evaluation splits carved by plant, capture campaign, or skill label

At a glance

Property	Value
Clips	259 across 14 process domains
Available sample footage	~100 hours
Compressed size	60.6 GiB, compressed from 809 GiB of source (14.7× mean)
Workers in sample	256
Clip length	≥5 minutes after curation
Modality	RGB video (+ AAC audio when present)
Viewpoint	Egocentric / first-person, wearable headset
Annotations	scene_segmentation JSON for every clip · human hand-pose demos
Format	H.264 MP4 (+faststart) in a VideoFolder layout
Split	Single training corpus (custom splits encouraged)
License	Private commercial license for authorized lab partners

Capture network

The Hub sample is a purchasable snapshot. Licensed programs draw on the full World Data Labs network under exclusive factory-to-lab agreements.

Capacity	Scale
Partner factories	135+ sites — U.S., Mexico, Canada, South America and more
Workers in contract	35,000+ front-line operators
Production workflows	1,000s of live manufacturing processes
Available now	~1,000 hours/week, capture hardware already deployed
Full scale (~2 mo ramp)	~100,000 hours/week as hardware is distributed

That network is the difference from the rest of the robotics data market, where capture concentrates in a handful of regions, workcells are ad-hoc, and the same clips are resold across labs. Here the footage comes from large automated plants with skilled operators on live lines, under exclusive contracts, streamed factory to lab.

Beyond RGB, licensed programs can add force/torque, depth, LiDAR, IMU, audio, and machine telemetry. Engagement is flexible — review samples first, per-hour pricing, raw egocentric video or video plus post-processed pose.

What is in the sample

Medical device and connector assembly · wire harness, electromechanical, and electronics module assembly · sterile device and polymer film packaging · plastics fittings assembly and plastic tubing handling · food and granola lines · bulk bagging and palletizing · warehouse pallet logistics · foam packaging warehouse · fulfillment document ops.

Folder	Clips	Compressed size (GiB)
bulk-bagging-palletizing	7	0.78
electromechanical-assembly	17	3.29
electronics-module-assembly	11	3.03
foam-packaging-warehouse	3	0.55
food-granola-line	6	0.84
fulfillment-document-ops	41	7.06
medical-connector-assembly	32	6.62
medical-device-assembly	37	7.63
plastic-tubing-handling	16	4.74
plastics-fittings-assembly	15	4.47
polymer-film-packaging	19	7.31
sterile-device-packaging	26	6.25
warehouse-pallet-logistics	16	5.44
wire-harness-assembly	13	2.56
Total	259	60.58

Filenames on the food and bagging lines carry task names ready for weak supervision: bulk-bagging-palletizing covers Bagging Palletizing, Bagging Sewing, and QC Sesame; food-granola-line covers Granola Filling, Packing, Tray, and Tray Cart. Other domains are keyed by process and timestamp.

Curation dropped low-quality and duplicate plant dumps along with any clip under five minutes, then renamed folders to industry / process domains. Additional sites roll onto the Hub as capture continues, so the table above is the current purchasable snapshot.

Scene segmentation

Every clip is cut into ordered subtasks: contiguous segments with a skill label, an outcome, the entities involved, and process narration. A six-minute video becomes a queryable list of steps.

Property	Value
Labeled clips	259 / 259
Scenes	3,911
Entities	2,081
Schema	annotations/scene_segmentation/schema.json (OpenAI-strict json_schema 1.0.0)
Sidecar path	annotations/scene_segmentation/<domain>/<stem>.json

Each record carries clip (path, domain, viewpoint, camera id, duration, resolution, fps), environment (setting, workcell, lighting, clutter, safety notes), entities (stable ids for people, hands, tools, workpieces, machines, PPE), scenes (start_time, end_time, skill, label, outcome, confidence, actor/object ids, narration), and a short VLA-oriented summary. Scenes also include a coarse bbox_xyxy_norm per entity — approximate region hints for weak supervision, not tracked detections.

Rendered straight from the shipped annotations — subtask list, per-step outcome, segment boundaries, and narration for the active step:

Because each step carries a skill, an outcome and exact boundaries, review becomes a filter instead of a scrub through footage:

where skill = "inspect" and outcome = "fail"
return clip_id, step, t0, t1

Domain	Example subtask run
Plastics fittings assembly	inspect → place → pick cycle, 5–10 s per step
Polymer film packaging	carry → place → walk → scan material movement
Wire harness assembly	long-horizon inspect and routing segments
Bulk bagging and palletizing	bag fill under hopper, then sew closed

Human hand pose

Hand-pose overlays expose the motion signal around manipulation — wrist orientation, finger articulation, grasp transitions, two-hand coordination — rendered on real first-person factory capture rather than staged tabletop footage. These are demonstration outputs for judging pose quality and retargeting potential, not dense ground-truth labels on every Hub clip.

Coils handling — fine manipulation and two-hand coordination around small industrial components. Open MP4.

Sesame quality control — hand tracking during inspection and material interaction on a food-processing line. Open MP4.

Files and format

videos/
  metadata.jsonl
  plastics-fittings-assembly/
    YYYYMMDD_HHMMSS.mp4
  electromechanical-assembly/
    ...
  food-granola-line/
    SC_GRANOLA_*.mp4
  bulk-bagging-palletizing/
    LS_BAGGING_*.mp4
annotations/
  scene_segmentation/
    schema.json
    plastics-fittings-assembly/
      YYYYMMDD_HHMMSS.json
    ...
assets/
  teaser.gif
  teaser.mp4
  teaser.jpg
  hand_pose/
    coils.gif
    coils.mp4
    sesame-qc.gif
    sesame-qc.mp4
  scene_segmentation/
    *_steps.jpg
metadata.json
README.md

• videos/<domain>/*.mp4: training clips grouped by process domain
• videos/metadata.jsonl: VideoFolder join table driving the Data Studio player, keyed by file_name = <domain>/<stem>.mp4, with domain, duration_seconds, has_scene_segmentation, n_scenes, n_entities, summary, and scene_segmentation
• annotations/scene_segmentation/: one JSON per clip plus the schema
• assets/: teaser, hand-pose demos, and the subtask-timeline stills shown above
• metadata.json: source_path, repo_path, sizes, optional compression_ratio / duration_seconds / uploaded_at

Headset model and native resolution vary by session, so delivered files are normalized for consistent training ingestion: H.264 (libx264) at CRF 28, medium preset, max width 1280 px with even auto height, AAC ~128 kbps where audio exists, MP4 with +faststart. This targets throughput into GPU training jobs; higher-bitrate masters are available under custom commercial agreements.

Usage

import json
import os
from pathlib import Path

from huggingface_hub import hf_hub_download

repo_id = "Worlddatalabs/egocentric-manufacturing"
token = os.environ["HF_TOKEN"]

video = hf_hub_download(
    repo_id,
    "videos/bulk-bagging-palletizing/LS_BAGGING_BAGGING+SEWING_1_CAM1.mp4",
    repo_type="dataset",
    token=token,
)

ann = json.loads(
    Path(
        hf_hub_download(
            repo_id,
            "annotations/scene_segmentation/bulk-bagging-palletizing/LS_BAGGING_BAGGING+SEWING_1_CAM1.json",
            repo_type="dataset",
            token=token,
        )
    ).read_text()
)
print(len(ann["scenes"]), "scenes;", ann["summary"][:160])

Load the viewer join table:

from datasets import load_dataset

ds = load_dataset("Worlddatalabs/egocentric-manufacturing", split="train")
row = ds[0]
print(row["domain"], row["n_scenes"], row["summary"])

License, access, and citation

Private dataset for licensed training use by frontier labs and robotics partners. Share your Hugging Face user ID for access to the sample; redistribution, public mirrors, or sublicensing require written agreement. Sales and licensing: human@worlddatalabs.com · worlddatalabs.com.

@misc{worlddatalabs_egocentric_manufacturing,
  title        = {Egocentric Manufacturing},
  author       = {World Data Labs},
  year         = {2026},
  howpublished = {Hugging Face Datasets},
  url          = {https://huggingface.co/datasets/Worlddatalabs/egocentric-manufacturing},
  note         = {Commercial egocentric manufacturing video for physical AI training}
}

metadata.json and the viewer table are refreshed from the live Hub tree by scripts/sync_egocentric_metadata.py; this card is regenerated by scripts/build_egocentric_readme.py.