Demonstrations from trained operators.
Imitation learning needs clean, consistent demonstrations. We supply recordings from trained operators, labeled episode by episode, with the rights to train on them.
Robotics teleoperation data, with proof.
Robot learning from demonstration depends on episodes that are consistent, successful and diverse in the right ways. Inexperienced operators produce hesitant trajectories, recoveries that look like noise, and inconsistent strategies for the same task. HUMXN works with trained teleoperation operators who record demonstrations to a written task protocol: initial conditions, object sets, success criteria, allowed strategies and how to handle failed attempts, so episodes are comparable across operators and sessions.
Recordings typically include joint states and actions, end-effector poses, gripper commands, and synchronized camera streams such as wrist and scene views, along with any force or tactile signals your setup produces. Each episode carries task and variation metadata, a success label, and annotations such as phase boundaries or failure reasons where your spec calls for them. Failed and recovery episodes can be kept and labeled rather than discarded, since some policies learn from them.
Every file is fingerprinted at intake, with frame hashes for video, so duplicate or re-submitted episodes are caught. Recordings that capture identifiable people are only offered for training with consent on file, and location metadata never reaches buyers. Episodes are delivered with a signed provenance record and listed by ID and hash on the receipt, so a policy checkpoint can be traced to the exact demonstrations it was trained on.
Built to your specification.
Every engagement starts from a written spec and a pilot batch. These are the most common requests we source for robotics teleoperation data.
Teleoperated demonstrations
Manipulation and mobile-manipulation episodes recorded by trained operators to your task protocol.
Synchronized sensor streams
Joint states, actions, poses and camera views aligned per episode, plus force or tactile signals where available.
Success and failure labels
Per-episode outcomes against written success criteria, with failure reasons where required.
Phase and event annotation
Sub-task boundaries, grasp events and contact moments marked by reviewers.
Egocentric video
Head- or chest-mounted recordings of people performing tasks, with consent and releases on file.
Task protocol design
Initial conditions, variations, object sets and success criteria written with your robotics team.
01Operator quality shows in the policy
Hesitant or inconsistent demonstrations teach hesitant, inconsistent behavior. Trained operators and a written protocol reduce that variance.
02Diversity on purpose
Variation in objects, positions and conditions should come from the protocol, not from operator mistakes.
03People in frame
Scene and egocentric cameras capture people and places. Consent and releases on file, and stripped location metadata, keep that usable.
04Traceable episodes
Hashes and signed provenance tie each policy version to the episodes behind it.
Robotics teleoperation data: asked often.
Which robot platforms and formats do you support?
Recording platform is agreed during scoping, based on your embodiment and where demonstrations need to be captured. Episodes can be delivered as WebDataset shards, Parquet tables or JSON Lines with linked media, with Croissant 1.0 metadata. If your stack expects a specific episode schema, we write it into the specification and validate deliveries against it.
How are teleoperation operators trained?
Operators are trained on the task protocol and the teleoperation interface before recording production data, and complete qualification episodes reviewed against the success criteria. During production, reviewers sample episodes for consistency and protocol adherence. Operators whose episodes drift from the protocol are retrained or removed from the project.
Do you keep failed episodes?
If your spec asks for them, yes. Failed and recovery episodes are labeled with the failure reason and phase rather than discarded, since some methods learn from them and they are useful for evaluation. If you only want successful demonstrations, failed attempts are excluded from delivery and the receipt lists only accepted episodes.
What affects the cost of teleoperation data?
Main factors are task complexity and episode length, the number of variations, the sensor setup, annotation depth such as phase boundaries, whether recording uses your hardware or ours, and exclusivity. Long-horizon, contact-rich tasks cost more per episode than short pick-and-place. We quote against the task protocol after a pilot run.
How do you handle people who appear in recordings?
Recordings that show identifiable people are only offered for training when their consent is on file, and releases are bound to the episode. Location and personal metadata are stripped before delivery. For egocentric collection, participants agree in advance to what is recorded and how it may be used.
Tell us what your model needs to learn.
Send a brief in five minutes. A data lead replies within one business day with questions and a first sourcing plan.