Egocentric data collection
for embodied AI
Real homes, kitchens, retail floors, warehouses and factories. Interviewed and trained human collectors record head-mounted video, hand motion and multimodal sensors in sync — delivered as training-ready ego datasets.
Before a robot can act,
it has to see how people do it
The first-person view naturally matches a robot's own sensor viewpoint and carries hand–object interaction, attention shifts and task context — exactly the data VLA and imitation learning need most, and the hardest to collect in-house.
- 1Viewpoint aligned with the embodiment
Head-mounted view ≈ robot head camera: low transfer loss, direct pre-training gains.
- 2Real, long-horizon, diverse
Actual homes and workplaces, not lab sets. Complete task chains including failures and recoveries.
- 3A fraction of teleoperation cost
No robot needs to be present — the human hand is the capture device. Tens of times more hours for the same budget.
What we collect
Specs are defined by your model inputs. Any combination of the modalities below, all on one aligned timestamp.
Egocentric video
Head-mounted RGB / stereo / depth video at 30–60 fps, fisheye or standard FOV, with intrinsics, extrinsics and distortion calibration.
Hand & body motion
21 hand keypoints, 6-DoF wrist pose, glove finger-joint angles; optional full-body IMU mocap and plantar pressure.
Multimodal sensors
Head IMU, eye tracking / gaze, ambient audio and speech, optional tactile and force sensing. Hardware clock sync.
From brief to delivery in five steps
You provide the task list and model requirements — we handle the rest. First pilot samples in as little as 7 days, then scale once specs are confirmed.
Define requirements
Task list, modalities, hours, scenes and collector profile
Plan & script
Collection SOP, action scripts and hardware setup
On-site collection
Executed in real scenes, daily uploads, live dashboard
QC & de-identification
Three-tier QC, segment-by-segment review, sensitive-data redaction
Deliver & iterate
Standard-format delivery, top-up collection based on model feedback
Measurable quality, auditable compliance
Every delivered clip traces back to its collector, device, script version and QC record.
Three-tier QC
- On-device checks: real-time alerts for frame rate / exposure / sync / occlusion; failed clips re-shot on site
- Segment-by-segment review by QC staff: task completion, viewpoint validity, adherence to the action script
- Sampling audit: frame-level checks per batch, with QC report and pass rate
- Acceptance metrics written into the contract: valid-hour ratio, sync error, clip pass rate
Privacy & compliance
- Informed consent signed by every collector and location owner, with purpose-bound authorization
- Faces, license plates, screens and ID documents de-identified automatically and by hand
- Encrypted transfer and storage; in-country delivery available, no cross-border transfer
- Raw data destroyed after the project as agreed, with a certificate of destruction
Diversity control
- Collector profile by spec: age, gender, handedness, height, occupational background
- Scene distribution by spec: floor plan, lighting, time of day, object categories and placement
- Same task repeated across people and scenes, including failure and recovery samples
Covering the work scenes robots meet every day
Real-environment data captured across a network of real locations. Common requests below — custom scenes available for your deployment.
Home living
Tidying, storage, cleaning, dishwashing and appliances
Retail shelves
Restocking, facing, scanning, picking and packing
Warehouse & logistics
Unboxing, sorting, palletizing, carrying and labeling
Factory floor
Insertion, fastening, cable routing, tool use
Office & desk work
Document handling, device operation, fine tabletop tasks
Elder care
Handing objects, assisted standing, daily care motions
Flexible hardware, matched to your model
You can supply the capture hardware and we run the collection, or we collect with our own ego capture rigs.
- Head-mounted RGB / stereo cameras (GoPro, Insta360, etc.)
- Smart-glasses multi-sensor devices (Aria-class)
- Depth cameras (RealSense, ZED, Orbbec)
- Mocap gloves and hand tracking
- Head / full-body IMU mocap
- Eye and gaze tracking
Three ways to work together
From a one-week pilot to long-term data supply, at your pace.
Pilot sample pack
Validate data specs and model gains at small scale and confirm direction fast.
- 10–50 valid hours
- Delivered in 7–14 days
- Includes collection SOP and QC report
Project-based collection
Fixed scope by task list and hours, delivered in batches with three-tier QC.
- Priced per valid hour
- Weekly incremental delivery
- Acceptance metrics in the contract
Long-term data supply
A dedicated collector team and committed capacity, delivering monthly and adjusting scripts as your model iterates.
- Dedicated collection team
- Monthly capacity commitment
- Priority scheduling and discounts
Frequently asked questions
Who owns the data? Will it be sold to others?
Data from project-based collection is exclusively owned by the client. Atisbo does not retain, reuse or re-license it. If you opt for the lower-cost shared-dataset licensing model, the terms are stated explicitly in the contract.
What is the minimum order?
Pilots start at 10 valid hours. We recommend a pilot to confirm specs before moving to project-based collection.
Can we use our own capture devices or data format?
Yes. We first validate device calibration and the sync setup, then output to your schema.
How do you make sure the recorded actions match our task definition?
We agree on action scripts and pass criteria during the requirements phase. Collectors are tested against the script before onboarding, on-site leads spot-check, and QC reviews every segment against the criteria. Failed segments are not billed.
How are privacy and compliance handled?
Collectors and location owners sign informed-consent forms; faces, license plates and screens are de-identified; data is stored encrypted, with in-country delivery and post-project destruction available.
Tell us what data
you need
Send us a task list and we'll reply with a collection plan and quote within 48 hours. Or book a 30-minute requirements call.