Egocentric (First-Person) Video Data Collection in Africa for Robotics and Embodied AI

In a phone repair stall in Lagos, a technician leans over a magnifying lamp and lifts a tiny connector with tweezers. Watch the same moment through a camera strapped to his forehead and you see exactly what a robot would need to see: two hands, a fragile part, the angle of approach, the small correction just before contact. That is egocentric video, and it is becoming one of the most wanted kinds of data in robotics and embodied AI.

What is egocentric video data?

Egocentric video is footage filmed from a camera worn by the person doing the task, usually on the head or chest. It shows the world the way the worker sees it: hands in the middle of the frame, tools and materials up close, attention moving from one thing to the next. Third-person (exocentric) video, by contrast, films the worker from the outside.

Research has leaned on a few big datasets. EPIC-KITCHENS-100 has 100 hours of unscripted kitchen footage from 45 kitchens. Ego4D grew that to 3,670 hours from 931 camera wearers in 9 countries, with parts that include gaze, stereo and 3D. Ego-Exo4D focused on skilled activity, pairing first- and third-person views with motion-sensor (IMU) data across 1,286 hours.

• What it captures: hand pose, grasp, tool use, object state changes, task order.
• What it adds: the viewpoint a robot's own head or wrist camera will have.
• What it needs alongside: timing, motion data and clear consent from everyone in shot.

Why robotics and embodied AI teams want it

Robots that work in the real world need to learn manipulation, and manipulation data is scarce and expensive to collect with robots alone. Human first-person video is one of the ways teams widen what their models have seen.

The field is moving towards general-purpose robot models trained on many kinds of data. Open X-Embodiment pooled data from 22 robots across 21 institutions, covering 527 skills, and showed that a model trained on the mix transferred skills between robots. π0, a vision-language-action model, was trained on a large, diverse dataset from several dexterous robot platforms. Both point the same way: more variety in tasks and settings helps.

1. Pretraining visual representations on how hands and objects move.
2. Action recognition and segmentation: knowing when one step ends and the next begins.
3. Hand-object interaction models for grasp and placement.
4. Evaluation sets drawn from settings the model has never seen.

Why skilled-trade footage from Africa matters

Because the real world is not a tidy lab kitchen. In a Lagos tailoring shop the machine may be treadle-driven, the light changes when the power goes, and the table is crowded with fabric and chalk. A mechanic at Ladipo works in open air with improvised tools. A phone repairer at Computer Village works under a magnifier on parts a few millimetres wide.

• Dexterity: sewing, micro-soldering, braiding and engine work are fine, two-handed tasks.
• Variety: different tools, materials, lighting and clutter from what most datasets show.
• Volume: these trades run all day, every day, so there is real work to record rather than staged demos.
• Policy fit: Nigeria's National AI Strategy, led by NCAIR, calls out the lack of local data as a barrier, so building it locally helps both sides.

How head-mounted phone capture works

A modern Android phone on a head strap is a capable egocentric camera. The phone sits flat on the forehead with the back camera facing out, tilted down 30 to 45 degrees so the hands sit in the middle of the picture. The worker looks at their work as normal and moves their head slowly.

• Video: 1920×1080 at 30 fps, H.264, about 4 GB per hour, from the main rear camera.
• Motion sensors: accelerometer and gyroscope (IMU) logged alongside the video at 50 to 100 Hz.
• Location and context: GPS every 30 seconds, a light estimate and battery level.
• Resilience: files roll every 10 minutes and uploads resume, so a crash or a dead network loses little.

Our standard capture is a single phone camera plus motion-sensor data. If you need stereo or depth, say so when you request a sample and we will scope the hardware with you.

Fraud controls: how you know a minute is real

Paying people per recorded minute invites cheating, so every control has to make faking harder than doing the work. These are the checks TrustAm Data runs on every clip:

1. App-only capture. Video can only be recorded in the app; nothing is picked from the gallery.
2. Signed capture manifest. Each session carries its ID, challenge code, device ID, app version, start and end time and a SHA-256 checksum per file, signed with a per-install key registered at login.
3. Challenge code. A random 3-digit code read aloud at the start, with both open hands shown to the camera.
4. Sensor match. The motion-sensor trace must look like a head-mounted worker. A phone lying still, or a still phone while the video moves, fails.
5. Duplicate checks. Frames are fingerprinted and compared across the whole collection, so re-uploaded or copied footage is caught.
6. Place check. GPS must fall inside the registered workplace's 150 m fence.
7. Person check. BVN or NIN plus a live selfie, one person per account, paid only to an account in the same name.
8. Human review. A trained reviewer approves or cuts each minute; contributors are paid only for approved minutes.

Consent and privacy in workplace video

Workplace video captures more than the worker, so consent has to cover the place and the people around it too. Each contributor records a consent line on video in their own language; the owner of each workplace consents before the first recording there; faces of customers and passers-by are blurred before delivery; and cash, ID cards and screens with personal details are never filmed. Contributors can withdraw at any time. This follows the Nigeria Data Protection Act 2023 and the NDPC GAID 2025 directive. Read the details on the trust and consent page.

How TrustAm Data collects egocentric video in Lagos

TrustAm Data runs a Lagos pilot with tailors, mechanics, phone repairers, cooks, braiders, welders, carpenters and market traders. The flow from workshop to your pipeline looks like this:

1. A TrustAm field agent registers each workplace in person, with the owner's consent, a shop-front photo and a map pin.
2. Each recorder passes an ID check, gets a free head strap and passes a short recording test.
3. Recorders film their normal work in batches opened for a buyer's order.
4. Automatic checks run, then a reviewer approves or cuts every minute.
5. You receive blurred MP4s, per-clip JSON, a manifest with consent IDs and a QA report, with LeRobot or RLDS on request. Label layers such as hand pose, action segments and captions can be added to order.

See first-person work video, browse the dataset library, read TrustAm Data for AI labs, or ask for custom collection. We also collect Nigerian speech data in English, Pidgin, Yoruba, Igbo and Hausa. For the commercial side, read Buying AI training data from Nigeria: a buyer's guide. If you are a worker who wants to record, start at trustamai.com/earn or read How to earn money from AI data in Nigeria.