THE EMBODIMENT RUNTIME FOR ROBOT FOUNDATION MODELS
Foundation models are the brain. Kyno is the body sense.
Drift, backlash, wear, latency. To the policy, a changed body is a distribution shift it never trained on.
A new body is a distribution shift.
Nothing alerted the policy. A miss measured in millimeters, paid in task failure.
A miss measured in millimeters.
DETECT
From what the robot just did, Kyno works out what is different about this body.
Read straight from what the robot just did.
ADAPT
The same intent, recompiled for the machine as it is today, not as it was on paper.
The same intent, recompiled for the body it has.
Same intent, corrected at runtime for the body it runs on.
Corrected at runtime, for the body it runs on.
LEARN
One unit's correction becomes the whole fleet's head start.
One unit's fix, the whole fleet's head start.
PROOF
Five bodies today, and this is just the beginning.
And this isn't just our lab. The same drift is already sitting in public fleet data. Nobody was reading the logs.
12 of 13 independent public SO-101 datasets carry a stable positive elbow offset. Different owners, different hardware, +1.2° to +3.9°. One (◆) is the LeRobot team's own dataset: +2.31°, drifting +0.014° per episode. The arm was drifting while its own dataset was being recorded.
Grasp success measured in simulation, 30 trials per condition: 30/30 on SO-101, 25/30 on xArm7. The remaining body schemas (stretched SO-101, Panda, G1-arm) are validated compile paths in sim. Field offsets computed from 13 public SO-101 datasets and independently reproducible.
WHY NOW
Robot foundation models are leaving the demo lab. The moment a policy runs on a fleet instead of one pampered robot, per-unit variation stops being a calibration chore and becomes a runtime problem. Cheap arms make it urgent. Expensive robots make it valuable. Humanoids make it existential.
Model vendors already feel it: Ai2's MolmoAct2 ships its SO-101 checkpoints "with calibration correction baked in." A correction frozen into weights is a fleet average: a static answer to a dynamic problem. Per-unit error is invisible in training data by construction, and it never stops. The fix has to live in the runtime loop. Whoever runs that loop owns the execution data that makes robot policies reliable on real bodies.
Have one robot that used to work?
Let Kyno find what changed.