Robots learn from data. But the moment that decides success, contact, is the one your data captures worst. Touch is the least trustworthy signal your model consumes.
We measure what happens at contact, so failures stop being mysteries.
A gripper slips. An insertion jams. A part gets crushed.
Nothing tells you the grip was two newtons short, or that the jam started at 0.3 seconds, or that the policy was slipping on the runs that "succeeded" too. So teams collect more episodes instead of fixing the failure.
Success rate tells you that it broke. It never tells you why.
We run your policy on your hardware and record what actually happened at contact: what the hand felt, what it did, when it jammed or slipped, what contact was on the part, and uncertainty. Every trial comes back with a contact-aware score, a timeline of contact events, and the evidence behind it.
That record plugs into the eval tooling you already run and fills the hole in your data stack where contact should be. No new controller, no new sensors. Each failure becomes a repeatable test, so the next checkpoint has to prove it improved.
Find the failure → measure it → reproduce it → retest it.
If your policy fails at contact and nobody can tell you why, send us the episode. A short description is enough to start.
Prefer email? info@tensiq.com. We'll use this information only to respond to your request. A short representative log or clip is plenty; please don't send anything confidential or export-controlled.