TrainThemAI/POV-Egocentric-Video-Robotics-FHD-Samples

POV Egocentric Video — Robotics FHD Samples

26 clips of first-person video of routine household 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.

---

📞 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

26 clips, ~27 GB total, spanning real-world Activities of Daily Living (ADL):

Domain	Clips
Cleaning	Bathroom Cleaning, Cleaning bathroom, organizing_bathroom, Tidying up kitchen, Tidying up living area, Dusting and organizing items, Cleaning keyboard keycaps, Cleaning and organizing items
Cooking & food prep	Cooking, Cooking cookies, Prepping meal, Making Coffee, Preparing Mate (Argentinian Drink)
Dishwashing	Washing dishes, Clean Dishes, washing_dishes
Organizing	Organizing clothes (×2), Organizing cutlery, Organizing cups and glasses, Sorting screws/nuts/nails/washers
Bedding	Making the bed (×2), tidying_the_bedroom
Other	Watering plants (×2)

Technical specifications

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, idle pauses kept under 10 seconds
Clip length	1–10 minutes (varies by task complexity)
Environments	Real homes across multiple locations, natural lighting
Total	26 clips, ~27 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.MP4 → Transit.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	residential
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:

• Behavioral cloning from human demonstrations
• VLA / WAM pretraining — observation-rich first-person video gives world-model training signal
• Fine-tuning π0, π1, RT-2, OpenVLA, GR00T on ADL task distributions
• Benchmarking egocentric perception, hand detection, and action recognition
• Quality reference when evaluating whether TrainThemAI's production pipeline fits your spec

How this compares to public alternatives:

Dataset	Scale	Focus	License	Production-extensible?
This sample	26 clips / ~27 GB	Household ADL, real-world clutter	MIT	✅ commercial pipeline
EPIC-Kitchens	~700 clips	Cooking only, academic	Custom (non-commercial)	❌
EgoExo4D	~5,000 hr	Multi-view skilled activities	Academic license	❌
Ego4D	~3,600 hr	Broad ego, low manipulation density	Academic license	❌

For research benchmarking, the above are excellent. For commercial-grade training data 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_pov_egocentric_2026,
  author = {TrainThemAI},
  title = {POV Egocentric Video --- Robotics FHD Samples},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/TrainThemAI/POV-Egocentric-Video-Robotics-FHD-Samples}
}

Contact

• Sales / contract data: hello@trainthemai.com
• Web: trainthemai.com
• LinkedIn: linkedin.com/company/trainthemai