Proprioception

A robot knows where its own joints are, precisely and at high frequency, without looking. That sense is cheap, exact and available every millisecond — unlike vision, which is slow, occluded and expensive — and a policy that ignores it is throwing away its most reliable input.

The practical difficulty is fusing signals at different rates: joint states at a kilohertz, cameras at thirty hertz, a language instruction once. Policies typically run a fast loop on proprioception with a slow loop on perception, which is the same hierarchy Hierarchical RL describes arriving from a different direction. Proprioception is also what survives when a camera is occluded, which is most of the time during contact.

With joint states at f sub _p  and images at f sub _v , a policy conditioned on both is limited to f sub _v  unless the architecture splits them — which is why the fast loop exists. During contact the visual signal is frequently occluded by the manipulator itself, so the information rate available to a vision-only policy drops exactly when the task is hardest.