Robot Demonstration Data
Capture demonstration data matched to your robot, environment, and evaluation goals, not generic datasets with mismatched action spaces.
Robot demonstration data is example episodes, observations paired with the actions taken, captured so a policy can learn your task. The data that works is matched to your robot, your environment, and your evaluation goals, not a generic dataset with a mismatched action space. Operant captures human, teleoperation, and egocentric demonstrations with clear episode boundaries, synchronized sensors, and audit-ready metadata.
What counts as a good demonstration
A good demonstration is recorded in your robot's action space, has clear start and end boundaries, is synchronized across sensors, and carries metadata describing task, operator, scene, and outcome. Demonstrations that miss any of these silently undermine downstream training.
Human vs. teleop vs. egocentric
- Human demonstration: a person performs the task; useful for reference and egocentric pipelines.
- Teleoperation: an operator drives the robot directly, captured via our teleoperation capture service.
- Egocentric: first-person capture of the behavior in a real environment.
We scope the right mix for your policy as part of imitation learning data collection.
Annotation options
Demonstrations can be delivered raw or with labels scoped to your schema, segmentation, phase tags, success annotations, or object references, applied consistently across the program.
Delivery formats
You receive synchronized sensor logs, calibration files, episode metadata, and optional labels in the formats your training stack expects, with documentation and provenance. For a concrete scenario, see dual-arm kitchen manipulation or the broader robotics training data guide.
FAQ
A good demonstration is recorded in your robot's action space, has clear start and end boundaries, is time-synchronized across sensors, and carries metadata describing the task, operator, scene, and outcome so it can be filtered and audited.
Human demonstrations show a person performing the task, teleoperation has an operator drive the robot directly, and egocentric capture records a first-person view. Each maps to a different learning setup; Operant scopes the right mix for your policy.
Deliverables include synchronized sensor logs, calibration files, episode metadata, and optional labels, exported in the formats your training pipeline expects, with documentation and provenance.
Scope your capture program
Book a discovery call to align on your stack and data requirements.