Clean demonstrations, at the volume learning needs.
Teleoperation is the fastest route to demonstration data and the fastest route to noisy datasets. Agentuor tracks quality per operator, segments sessions into episodes, and filters demonstrations before they reach training.
The challenge
Demonstration quality varies by operator, by hour, and by task familiarity. A hesitant grasp, a corrective wiggle, or an aborted attempt all end up in the training set unless someone catches them. At scale, nobody can watch every session, so the model learns the hesitation. Programs also lack the feedback loop that would let operators improve: they rarely learn which of their demonstrations were kept.
Built for this data
Session segmentation
Continuous sessions cut into episodes at task boundaries, with attempts and aborts detected.
Operator quality tracking
Smoothness, completion, and correction metrics per operator and per task.
Demonstration ranking
Agents score demonstrations for cleanliness and recommend which to include, exclude, or review.
Intervention marking
Autonomy-to-teleop handovers and corrective interventions labeled as events.
Operator feedback
Operators see which demonstrations were kept and why, closing the loop.
Provenance
Every inclusion or exclusion decision is recorded with its reason.
Data and labels
- Operator commands and robot state at control rate
- Synchronized video from robot and rig cameras
- Episode, attempt, and abort boundaries
- Intervention and handover events
- Operator identity, task, and scene metadata
Typical workflow
Record sessions
Rigs stream to Agentuor with operator and task metadata.
Segment and score
Episodes extracted; cleanliness scores proposed per demonstration.
Confirm and label
Reviewers confirm inclusions, exclusions, and outcomes.
Export training sets
Filtered demonstration sets exported with full provenance.
Outcomes teams work toward
Common questions
Which teleop rigs are supported?
Any rig that logs commands and state; SDKs cover common middleware.
Can operators be anonymized?
Yes. Operator identifiers can be pseudonymized while preserving per-operator metrics.
Does filtering happen automatically?
Agents recommend; reviewers confirm. No demonstration is excluded without a decision on record.
Show us an episode. We'll show you the gaps.
Start with a walkthrough on your own data.