Robots learn to open doors from data like this
Your work email is the key. Robots have to do it the hard way.
Behind the door: our full capabilities and the sample request form. Episodes land in your inbox within a business day; a human may follow up - one email, not a drip campaign.
Physical AI Data Services
Human demonstration and teleoperation data for robot learning - collected across real US homes and our Los Angeles facility, delivered post-QA in your training format. For robotics teams and research labs, from a first sample pack to fleet-scale programs.
Proven at fleet scale
We deployed and fleet-managed the ~2,000-headset capture fleet behind one of the largest home robot-learning programs run to date - 10M+ episodes collected across real US homes. Kitting, logistics, remote fleet management, QA. The same operation now runs your program.
What we can collect
Egocentric
Head-worn camera + optional wrist cams. Real tasks, real homes.
UMI / wearable glove
Ego video + hand-worn capture: wrist pose, finger tracking.
Dexterous hand
Station gloves, externally-tracked ground truth, full articulation.
Robot teleop
Operators driving arms + dexterous hands. Embodiment + force.
Data channels by mode
| Ego | UMI / Glove | Dexterous | Teleop | |
|---|---|---|---|---|
| Ego video | ✓ | ✓ | ✓ | ✓ |
| 3D camera tracking | ✓ | ✓ | ✓ | ✓ |
| Wrist camera | ✓ | ✓ | opt | ✓ |
| 3D hand tracking | coarse | ✓ | ✓ full | ✓ |
| Depth | opt | ✓ | ✓ | ✓ |
| Force / contact | — | — | opt | ✓ |
How each mode runs and scales
| Ego | UMI / Glove | Dexterous | Teleop | |
|---|---|---|---|---|
| Deployment | In-home | In-home | LA facility | LA facility |
| Onboarding | Ramp to 50/wk | Ramp to 50/wk | Per seat | Per station |
| Hrs / worker | 8–12 /wk | 5–10 /wk | 100–130 /mo | ~100 /mo |
| Data character | Volume | Breadth + pose | Ground truth | Embodiment |
Capacity and ramp - what your program can draw on
Today
4–6 hrs/day
Active cohort
Ramp (months 1–2)
10–30 kits/wk
Onboarding while we scale up
At scale (month 3+)
50 kits/wk
Sustained onboarding
Full network
2,000+
Collectors across our distributed network at full build-out
No-hardware deployment - activate existing VR owners
For egocentric collection we can activate existing VR headset owners - no hardware shipment, onboarding is software-only. ~500 collectors in month one, ~1,000 by month three, drawing on our decade of VR community reach. BYO collectors can be upgraded to glove-mode with a single small shipment (glove only, no headset). All quoted hours are post-QA delivered hours.
Program details
| Bespoke hours | Yes - immediate for distributed modes; dexterous-hand and teleop at pilot scale, expanding with LA facility (Q4 2026). Custom rigs, “your hand / glove / arm on our operators,” delivery in HDF5, RLDS/LeRobot, or buyer spec. |
| Collector network | 40–120 kit-based collectors (ramping) plus ~500 BYO-headset egocentric collectors ready to activate |
| Environments | Real homes nationwide - kitchen, laundry, living, bedroom, garage. Staged kitchen / bedroom / warehouse sets in LA. Office and light-industrial on request. |
| Tasks | Random and scripted object manipulation, pick-and-place, articulated objects, folding and deformables, tool use, bimanual, household routines, long-horizon chores. |
| Notes | We source and build capture hardware at distributor pricing (MANUS, SenseGlove, Xsens, dexterous hands, collaborative arms). We package, QA, and ship kits to end users across the US and fleet-manage remotely. Already buying gloves, suits, or arms from us? We can run collection, validation, or a managed program on the same hardware - one team for both. |
How a program starts
Samples
Request below. Sample episodes and schema docs arrive within one business day, in the format your program would receive.
Pilot
A scoped task set on the modes you pick. You evaluate real delivered episodes against your training stack before committing.
Scale
Ramp per the schedule above - kits in homes, facility seats, or both - with recurring post-QA delivery.
Request sample data
Pick the capture modes you want to evaluate and we’ll send matching sample episodes - post-QA delivered data from the same pipeline your program would run on, in HDF5 and LeRobot format with schema docs.
Knoxlabs · Glendale, CA · knoxlabs.com