What Haptica does, exactly.
We measure load presence. Whether the gripper is carrying load, with calibrated confidence, and abstention where the scene doesn't support a read. Magnitude needs an instrument: we tested whether anyone can judge force from video, and people don't do it reliably.
We find the spans where failures tend to happen. Onset and release neighborhoods, sustained contact, load intervals. Jams, slips, misalignment, and failed attempts are outcomes: Haptica preserves outcome labels where the source dataset has them, attributed to the source, and does not infer outcomes from video.
We work offline, on recordings. The labeling model runs after the fact, where a slow inference costs nothing. Your control loop stays yours.
We label what the camera supports. Every event carries confidence and a score for how clearly the camera saw the moment. When the scene doesn't support a label, the system declines, and the abstention stays in the record.
We publish results with their scope. Simulation results carry the word simulation every time. The real-robot result isn't claimed until it exists, and it publishes pass or fail.
Common questions
What does Haptica label?
Contact, onset, release, persistence, load presence.
Do I need force sensors?
Not for a released pack. The labeling method is built and checked on datasets that have them. For your own data, we agree up front on how label quality will be checked.
What do I receive?
Training track, event record, splits, quality and coverage measurements, provenance, a scorer, an integration recipe.
Where does my data go?
Pilots can run on your infrastructure. What you send is retained for your engagement and deleted when you leave; retention terms are in the agreement.
How do I try it?
Start with the pilot brief.