
October 14, 2026 · 12 min read
Robot Foundation Models Are Not Black Boxes
A robot foundation model takes an open-ended instruction, looks at what the robot sees, and outputs the actions that move the machine. That generality replaced the explicit kinematics, dynamics, and control of the classical stack, and the easy assumption is that the trade bought you an unreadable model. The first framework for interpreting and steering these models through their own internal representations, out of Claire Tomlin's lab at Berkeley, shows otherwise: project the feed-forward activations onto the model's own vocabulary, cluster them by concept, and you can turn behavior up or down at inference with no retraining. Here is what that means for the questions you should be putting to whoever sells you the model.
Read more →



