Contact
Whether physical contact is present.
Add physical state to your robot data.
Haptica labels contact, onset, release, persistence, and load presence in robot video. The labels join to your episode IDs as a physical-state sequence and a physical-event record.
The state that decides the task isn't in your labels.
Video records motion. It rarely records when contact starts, holds, or releases. You cannot add a force sensor to last year's recordings. You can add physical-event labels.
Instrument-grounded labels.
On instrumented robot data, Haptica contact labels track force-torque truth at 0.89 window-level AUROC, measured in-distribution. For each dataset, we report held-out quality, calibration, and abstentions.
The five signals.
Onset
When contact begins.
Release
When contact ends.
Persistence
How long contact remains active.
Load presence
Whether the gripper is carrying load. Presence, not force magnitude.
Built for training and inspection.
The labels come in two forms.
Physical-state sequence
A per-timestep sequence aligned to observations, actions, and language.
Physical-event record
Timestamped intervals with confidence, abstention, provenance, and evidence.
Join by episode ID.
source episode + Haptica labels = training-ready physical supervision- Join the labels to source episodes.
- Add the physical-state sequence to observations and actions.
- Inspect and filter with the physical-event record.
- Run the included baseline comparison.
Delivered as MCAP and LeRobot-compatible outputs. Our open Physical Events Schema publishes in August.
An ID-keyed overlay attaches the labels without repackaging the source video.
Labels for the datasets you train on.
DROID labels
Broad multi-scene manipulation.
Scheduled, license-cleared
BridgeData V2 labels
The corpus many teams already fine-tune on.
Scheduled, license-cleared
RoboMIND labels
Multi-embodiment coverage.
Scheduled, license-cleared
Instrument-grounded labels use datasets with force-torque traces to measure label quality.
Or run a pilot on the data you own.
Choose one task and one baseline. We agree in writing on what success means before labeling starts, then compare on the same data and compute.