Following the work on my M.Sc. (the Reactive Accompanist) in 1991 I realized that Behavior-Based AI, at least of that time, worked better as a model of established, expert behavior than of novice behavior. It gives no indication of how such behavior modules might be developed within the lifetime of an agent. Consequently, I became interested in the "consolidation" of new, skilled behavior, and began studying both natural and AI learning systems.
Behavior-based AI was originally meant to be purely reactive, which didn't allow for learning. If there's no variable state, where do you store what you learn? Although this problem was quickly recognized, the current learning research in autonomous agents often makes the opposite mistake: trying to use monolothic learning systems to develop the entire action-selection system for an agent. This loses the advantages of BBAI. My approach is that learning, like the rest of AI, should be modular and constrained to task, and an integrated part of perception and action.
Under my BOD methodology, most learning occurs within Behaviors, which are specifically designed for that process. There is also the potential to learn new plans, or to learn entire new behaviors, but I haven't been exploring those areas myself. The kind of learning I have done within behaviors is everything from very short term perceptual memory for disambiguating sensor inputs to long-term episodic and consolidated navigation memory.
For my publications on these topics, see my publications You may also want to see other people's work on consolidation and learning; see my related research web page for some of that.