Measured contact records
Label quality, event timing and robot training results.
Contact label quality
Validation method
REASSEMBLE: 83 unseen test recordings checked against wrist force-torque measurements; operating point selected on 61 development recordings. Contact precision uses a 95% interval.
At its shipped operating point, the impact channel matched 17 of 20 strikes on an axe clip within 500 ms: 187 ms median error and one unmatched record. Five non-impact clips returned one impact record.
REASSEMBLE label precision and coverage.
Robot training
Pre-registered robomimic Can simulation, with the robot's own contact signal at execution.
Study details
Haptera contact labels raised success from 40.0% to 49.8% (+9.8 points; 95% interval 2.0 to 17.7). Shifted labels reached 36.1%. Each arm ran 30 seeds with 30 rollouts per seed: 900 rollouts per arm.
Hard-case results
On the 63 pre-registered hard cases that failed all three baseline attempts, our labels recovered 39 (62%); the sealed simulator-sensor reference 40 (63%); the width-matched sensor reference trained alongside ours 44 (70%); shifted labels 23 (37%). Recovery showed no detectable difference between ours and either sensor reference. Correctly timed labels significantly outperformed shifted labels.
Simulation, robomimic Can. Labeled policies used the robot's own contact signal at execution.
Task success and hard-case recovery.
Contact at execution
Control comparison
The same contact channel with unlabeled training reached 32.4%, versus 42.3% for its control (interval minus 21.7 to plus 1.6 points).
Training with labels beat the channel-only policy by 18.6 points (interval plus 9.6 to plus 27.7).
The same contact channel at execution.
Labels teach the model.
Adapting the video model on force-labeled recordings cut missed onsets by a third at the same false-alarm rate, on recordings it had never seen. Labels teach the model, not only the policy.
Matched false-alarm rate, 83 unseen recordings.
Contact records
Haptera records contact and impacts from robot video, with timestamps, confidence and source evidence.
Contact labels alongside frames and robot logs.
The record
Versioned event files on your recording timeline.
- Contact state and impact times, with confidence and coverage.
- Contact starts: automatic timestamps with joint-effort logs; reviewed intervals from video.
- Recording IDs, model version, calibration and event evidence.
Inspect events.
Check each event against the source footage and share the evidence.
Labels for learning
Choose training moments and add contact labels on the same timeline.
- contact present
- abstained
- onset interval
- impact
Contact and impacts on the recording timeline.
Recording inputs
- 30 fps video of the interaction.
- Robot-state logs, if recorded. Joint efforts enable automatic contact-start timing at your chosen false-alarm rate within the measured held-out range.
- Name the interaction before setup: gripper to object or tool to workpiece, for example.
What each input adds.
Setup validation
Validation details
Before delivery, we validate your robot, camera, task and timeline against robot force, torque or joint signals, or a short capture with temporary instruments. Completed recordings are processed in batches.
Camera moves, task changes or measured drift trigger fresh validation.
- Your recordingsCamera, task and recording timeline.
- Check against physical sensorsRobot force, torque or joint-effort data, or a short capture with a reference sensor.
- Records matched to your recordingsVersioned files in your layouts, with a quality and label-coverage report.
- Setup validationConfirmed before delivery.
- Camera moves, task changes or measured driftFresh validation.
From setup to delivered records.