Teleoperation demonstrations
Human-controlled robot demonstrations captured at the task level.
Data built around the moments robots struggle with.
Demonstrations establish the task. Takeovers show where autonomy fails. Recovery turns that failure into a useful training signal.
Egocentric and wearable human demos
Human motion and manipulation captured from the operator's perspective.
Takeovers and recovery data
Human interventions when autonomous behavior fails.
Success / failure labels
Structured outcome labels tied to the episode.
One episode. Every signal aligned.
This schema is illustrative and adapts to the receiving training stack.
episode_id string | arraytask_id string | arrayrobot_id string | arrayoperator_id string | arraytimestamp float64camera_streams string | arrayjoint_states string | arrayaction_trace string | arraysuccess booleanfailure_type string | arraytakeover_start string | arraytakeover_end string | arrayrecovery_segment string | arrayqa_status string | array
Start small. Scale the loop.
The initial collection is a fixed operating target. The path beyond it is illustrative and agreed after review.
500demonstrations
20–40hours of manipulation data
50–100episodes in first review batch
10business days target delivery
500→2,000→10,000+
Start with the task.
Describe your robot and receiving stack to shape a relevant demonstration and takeover package.
How the loop works.
Scope
Define the task, embodiment, success criteria and data requirements.
Capture
Record demonstrations, interventions and recovery behavior.
QA
Review synchronization, trajectory integrity, labels and episode quality.
Deliver
Package data in the format the training stack already reads.
Close the loop
Feed recovery and failure data back into the learning process.