Physical interaction data for robotics foundation models.
Teleoperation, egocentric capture and custom environments: the demonstration data that trains VLA models and imitation learning policies. This data does not exist online. It is produced by trained operators, with hardware, in real environments under controlled protocol.
Estimate your dataset.
Billing is per hour of manipulation video collected and validated at QA. Present hours, rejected footage, training and reconfiguration time are never charged.
Estimate only. Final quote may vary. Prices exclude VAT, in euros. Not a quote or a contractual offer.
/// Public estimate for collected video hours only. UMI kits, launch packages, custom environments and exclusivity licences are scoped separately.
From data to a robot that acts, one pipeline.
Data collection
Proprietary, real-world interaction data captured at scale, structured, annotated and quality-scored into model-ready datasets.
Model training
Those demonstrations train Vision-Language-Action and imitation-learning models that turn perception and an instruction into action.
On-robot deployment
Trained policies run on the robot itself, on onboard compute, with standard formats and no cloud dependency for execution.
Robot teleoperation
Operators teleoperate the robot on real sites, and every run captures fresh demonstrations that feed back into step one.
What a demonstration actually looks like.
One process, tuned to your models.
Fine manipulation
Grasping, insertion and assembly gestures at high precision.
End-effector & hand trajectories, contact forces, multi-view sync
Manipulation & grasping policies
Sub-millimeter
Logistics & industrial environments
Handling and movement in real operational settings.
Body & payload motion in situ, spatial context
Warehouse robotics & world models
Centimeter, real-scene
Complex anthropomorphic gestures
Full-body and dexterous human motion, faithfully modelled.
Kinematic chains, degrees of freedom, timing
Humanoids & imitation learning
High-DOF, temporally exact
Captured, not scraped.
Data is produced by us, with a clean provenance chain and usage rights defined per engagement. Every episode is traceable end to end.
Full provenance
Every episode carries full provenance, so you know exactly who captured what, where and how.
Consent & NDA
Individual consent and NDA coverage for every operator involved in a capture.
Encrypted chain of custody
An encrypted chain of custody end to end, from capture to the delivered training file.
AI Act aligned
European contractual standards, aligned with AI Act traceability requirements.
Straight answers.
Trained operators control robotic arms in real time, in a leader-follower configuration, to execute manipulation tasks. Every session records the full action-observation stream, with synchronised RGB, wrist pose and gripper state, the demonstration format that imitation learning and VLA models are trained on.
First-person recording of humans performing real tasks with their own hands, captured with UMI-class rigs. It is the highest-fidelity source of natural manipulation behaviour, with no robot-morphology bias and no operator learning curve, at a fraction of the cost of robot-based collection.
LeRobot, RLDS and HDF5, or a client's internal schema. Each episode ships with synchronised streams, 6-DoF pose and gripper channels, timestamped action trajectories and structured metadata.
Yes. We build environments to specification (kitchens, bathrooms, workshop benches, retail shelving, warehouse picking stations, navigation circuits) and can reconstruct a client's own operating environment on our sites. Typical lead time is 4 to 6 weeks.
We run 100+ trained operators in production today at our dedicated Madagascar data factory, launched January 2026. Daily capture capacity is ~800 hours a day today, up from a standing start in January as headcount ramped to 100 operators, with 100,000+ hours collected to date. We're scaling toward a 13,000-hour daily target with 1,000 operators. Monthly volume is scoped per engagement.
Usage rights are defined contractually per engagement. Full provenance and individual consent are delivered with every batch, and nothing captured under exclusivity is resold.
Pricing is per engagement, driven by modality, volume and environment complexity. Request a data sample and we will scope it with you.
Request a data sample.
Tell us the modality, format and environment you need. We will return a representative sample and scope volume and delivery with you.