The Cerebellum of Physical AI: The Missing Layer That No Training Regime Can Replace

The Physical AI industry is investing heavily in approximating the cerebral cortex, the seat of reasoning, planning, and deliberate action. What it has not yet built is anything resembling a cerebellum. This article argues that the absence of a cerebellar equivalent in Physical AI systems represents a structural limitation that no training regime, however large or diverse, can address. Drawing on neuroscience, the article distinguishes between states, the fast-changing dynamic variables of physical systems, and parameters, which govern the underlying dynamics themselves. The cerebellum continuously estimates both in real time, providing the cortex with the physical grounding it requires to act reliably. Current Physical AI systems have no equivalent capability. The article further examines why conventional estimation tools, including Kalman filter variants, cannot solve the underconstrained nonlinear parameter estimation problem that characterizes real-world physical systems in deployment. The sim-to-real gap is reframed not as a data and fidelity problem but as a parameter estimation gap, a structural limitation that data alone cannot close. The article concludes with an open question: what would a genuine cerebellar layer for Physical AI actually require, and why does building it matter for the safe deployment of autonomous systems alongside humans?