Dataset Viewer

The dataset viewer should be available soon. Please retry later.

Industrial & Workplace Egocentric Video — FHD Samples

21 clips of first-person video of real industrial and workplace tasks, captured with a head-mounted smartphone. Released by TrainThemAI for training Vision-Language-Action (VLA) models, World Action Models (WAM), and humanoid manipulation policies — π0, π1, OpenVLA, RT-2, GR00T, Cosmos, DreamZero.

Fully rights-cleared, MIT-licensed, and representative of our production capture pipeline.

Companion to POV Egocentric Video — Robotics FHD Samples (residential / household tasks).

📞 Production-scale data — talk to us

We collect egocentric video at scale for embodied-AI teams.

  • 500+ active operators across Latin America and the Philippines (live as of May 2026)
  • Custom activity coverage — household, workplace (manufacturing, retail, hospitality, construction), or specialty domains
  • Per-project QC against client-specified rejection criteria
  • Typical engagement: 100–5,000 hours, 3–12 week delivery windows
  • Ego + wearable hardware dataset coming June 2026 — first-person video paired with hand pose and wrist trajectory tracking, for action-labeled data at ~1/10 the cost of robot teleoperation

📧 hello@trainthemai.com — we respond within one business day.
🌐 trainthemai.com

What's in the sample

21 clips, ~5.1 GB total, spanning real industrial and workplace environments:

Domain Clips
Manufacturing & factory work Lathe / handwheel spindle, Copper factory, Metal fabricator factory, Packaging factory, Textile factory, Ceramics, Woodworking, Factory (generic)
Retail, storefronts & food service Retail, Storefronts, Pharmacy, Restaurants & cafes, Barista
Personal services Hair / nail / lash services (×2), Tailor, Laundromat
Construction & heavy equipment Construction, Heavy equipment
Field & outdoor Farms, Transit

Technical specifications

Spec Value
Resolution 1080p (1920×1080)
Frame rate 30 fps
Codecs H.264 / HEVC video, AAC audio
Camera Smartphone with ultrawide (0.5×) lens
Mount Head strap at forehead or eye level, angled ~45° downward
Face Never on-camera by design
Hands in frame >90% of recording duration
Action density Continuous manipulation in operational settings, idle pauses kept under 10 seconds
Clip length 30 seconds – 6 minutes (varies by task)
Environments Real factories, workshops, retail floors, service businesses — operator-owned or operator-affiliated worksites, natural lighting
Total 21 clips, ~5.1 GB, MIT license

Per-clip metadata (JSON sidecars)

Every clip ships with a companion JSON sidecar carrying camera and capture metadata, named to match the video — e.g. Transit.MP4Transit.json, fetched from the same path. A combined metadata_manifest.json at the repo root indexes every clip.

Each sidecar contains:

Field Notes
session_uuid Stable per-clip identifier
environment_type commercial
country ISO code where embedded; unspecified otherwise (these clips carry no GPS)
camera_model Detected device class — GoPro HERO12 Black where onboard telemetry is present, smartphone (ultrawide 0.5×) otherwise
focal_length Physical focal length in mm
distortion_coefficients OpenCV radial/tangential [k1, k2, p1, p2, k3]
capture resolution, frame rate, codec, lens
imu_available / pose_available true for GoPro clips (onboard accelerometer + gyroscope and orientation/gravity telemetry); false for smartphone clips
calibration_status reference_nominal — intrinsics are reference values for the detected device class. Per-unit checkerboard calibration is available on request for production engagements.

These fields follow common egocentric-data intake requirements, so the samples can be evaluated directly against a production spec.

Why egocentric for embodied AI

The first-person, head-mounted perspective closely matches a humanoid robot's head-camera viewpoint, which makes this format especially well-suited for:

  • Industrial cobot / mobile manipulator policies — assembly, packaging, sorting, machine tending, material handling
  • Service-robotics fine-tuning — retail floor tasks, food service, personal-services manipulation
  • VLA / WAM pretraining on workplace task distributions, not just home or lab
  • Behavioral cloning for vocational skills — lathe operation, woodworking, textile handling, tailoring
  • Benchmarking workplace-egocentric perception, tool-use recognition, and procedural action segmentation

How this compares to public alternatives for industrial / workplace egocentric video:

Dataset Scale Focus License Production-extensible?
This sample 21 clips / ~5.1 GB Real factories, workshops, services MIT ✅ commercial pipeline
HoloAssist (Microsoft) ~166 hr Industrial maintenance / assistance Research
Assembly101 ~513 hr Toy assembly only Academic
Ego4D ~3,600 hr Broad ego, low industrial density Academic license
EgoProceL ~62 hr Procedural tasks, academic Academic

For research benchmarking, the above are useful. For commercial-grade training data in real industrial and workplace settings at the scale and spec you need, that's where TrainThemAI comes in.

License

MIT — free for any use including commercial, research, redistribution, and model training. Attribution to TrainThemAI appreciated but not required.

Citation

@misc{trainthemai_industrial_workplace_egocentric_2026,
  author = {TrainThemAI},
  title = {Industrial \& Workplace Egocentric Video --- FHD Samples},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/TrainThemAI/Industrial-Workplace-Egocentric-FHD-Samples}
}

Contact

Downloads last month
66