Robot learning data, collected at scale and checked at the source.
MobileEgo and the Simulation Data Platform are designed to support distributed collection of human demonstrations and robot training trajectories.
Two platforms, one library.
Real-world and simulated data share one task vocabulary, one review process and one versioned library.
- Phone captureMobileEgoStatus: Pilot
- SimulationSimulation Data PlatformStatus: Pilot
Every contribution
Checked at the source
- Verified devices and signed captures
- Independent replay of every episode
- Review, duplicates refused
- Faces masked on the phone
Delivered as
A versioned dataset
- LeRobot format
- Lineage to every source
- Quality report
MobileEgo
Status: PilotFirst-person video of people doing everyday tasks, recorded on their own phones. Each phone is enrolled and verified, captures are signed on the device, faces are masked on the phone before upload, and every capture is reviewed before it is accepted.
- Egocentric video
- Verified devices
- On-device masking
- Human review
Simulation Data Platform
Status: PilotContributors steer simulated robots from their phones to complete tasks in RoboCasa kitchens, show where policies fail and demonstrate the correction. Every episode is recorded with full state and actions and replayed independently before it counts.
- Teleoperation
- RoboCasa
- Replay-checked
- Corrections
Checked before it counts.
Every source has its own checks. Nothing reaches a dataset until it passes them.
- 01
Verified devices
Phones are enrolled with hardware attestation, and captures are signed on the device that made them.
- 02
Independent replay
Simulation episodes are replayed from their recorded actions; any that do not reproduce are rejected.
- 03
Review
Quality, completeness and task match are checked, and duplicates and copied footage are refused.
- 04
Privacy
Faces are masked before data leaves the phone, and contributors consent to training use first.
Data for each stage of training.
Pre-training data
Successful demonstrations and expert trajectories across the tasks, objects and scenes you specify.
Post-training data
Failures, interventions and corrections, each paired with the model checkpoint and scene it came from.
How a data project runs.
- 01
Scope
You tell us the model, robot, tasks and formats. We agree acceptance criteria up front.
- 02
Pilot
A small batch, delivered quickly, to test in your own training pipeline.
- 03
Campaign
Collection opens to the contributor network, aimed at exactly what you need.
- 04
Verify
Every contribution passes its source checks before it is accepted.
- 05
Deliver
Versioned datasets in LeRobot format, with lineage back to each source and a quality report.
Delivery and rights
- Formats
- LeRobot today; other formats are added per project when your pipeline needs them.
- Provenance
- Each delivered episode links back to its source contribution, its version and the checks it passed.
- Rights
- You receive the usage rights agreed in your contract. Custom data is not resold unless the contract says so.
Build with Robotensor
We collaborate with robotics companies, research labs, and industrial partners on robot learning, simulation, data collection, and evaluation.
Contact us to discuss your robotics applications, data requirements, or research challenges.