Kepler v0.1 trains robot vision, touch, and pose into one shared world model

freelerobot · x · 2026-07-23

Menlo Research introduces Kepler v0.1, an open research experiment where a robot’s vision, body position, and touch are trained together.
The core idea is that the senses can predict one another: from camera input alone, the robot can infer some information that previously required touch.
One encoder works across four different robots, and disagreement between sensory streams becomes a safety signal.
The team says the project is being trained in public, with failures published alongside results and the whole pipeline shared openly.

More from Embodied