Brian Salomaki, Dongkyu Choi, Negin Nejati, and Pat Langley
In this paper, we review teleoreactive logic programs, a formalism for expressing goal-directed reactive con- trollers for physical agents. Although these programs previously had to be entered by hand or learned from problem solving, here we present a new way to acquire them from observation. The learning system observes the traces of primitive skills taken from an expert work- ing on a known problem and learns new higher level skills based on this information. We explain the algo- rithm used for learning by observation and present ini- tial results in two domains. Finally, we review related work and discuss directions for future research.