Learning Robots

Associative memory is central, we associate different experiences, knowledge , facts with a certain abstract symbol.

Intelligence is embodied.

The cognitive cycle: adapt (reinforcement) – detect (sensors) – classify (response)

Models of learning need to be able to make the transition from continuous signals coming from the world to internal symbols.

One possibility is HMMs, these can be nested recursively. The idea is to use these nested models to build increasing complex symbolic representation.

Sensors: audio, video (contours, edges, histogram, …), proprioception, haptics

Bootstrapping with hard-wired behaviors (instinct): irritability, audio/video preferences, threats, hunger, detection of boundaries, reward system

Robot can be taught to perform certain actions (waving, draw a circle with its hand) when instructed to, much like how one would teach a dog.