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. Overlays join on your episode IDs: a per-timestep training track plus an auditable event record.
The state that decides the task isn't in your labels.
Manipulation policies fail at contact. The motion looks right, then the grasp misses or the part slips. A correct grasp looks identical to one crushing the part, and the deciding state is absent from every dataset's labels. You can't retrofit a force sensor onto last year's recordings. You can retrofit labels.
Grounded against physical instruments.
On instrumented robot data, Haptica contact labels track force-torque truth at 0.89 window-level AUROC, measured in-distribution. Every pack reports its own 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.
Every Haptica pack contains two connected outputs.
Policy-training sequence
A per-timestep physical-state track aligned to observations, actions, and language. Add it to the training records your policy already consumes.
Physical-event record
A timestamped log with confidence, abstention, provenance, and supporting evidence. Use it to inspect the supervision, build subsets, and evaluate models.
The overlay joins by episode ID.
source episode + Haptica overlay = training-ready physical supervision- Join on episode ID.
- Add the state track to observations and actions.
- Inspect and filter with the event record.
- Run the included matched-comparison recipe.
Works with the physical-events schema, MCAP, and LeRobot-compatible pipelines.
Supervision for datasets already in your stack.
DROID
Broad multi-scene manipulation.
Scheduled, license-cleared
BridgeData V2
The corpus many teams already fine-tune on.
Scheduled, license-cleared
RoboMIND
Multi-embodiment coverage.
Scheduled, license-cleared
Sensor-grounded packs: small corpora where labels are checked against the robot's own force sensors. Rights review underway.
Or start with the data you own.
One task, one frozen baseline, success criterion agreed before labeling. Then a matched evaluation on your robot, same data, same compute.