Abstract:
Robots assisting humans in complex domains often have to reason with different descriptions of incomplete domain knowledge. It is difficult to equip such robots with comprehensive knowledge about the domain and axioms governing the domain dynamics. This paper presents an architecture that enables interactive and cumulative discovery of axioms governing the action capabilities of the agent and the preconditions and effects of actions. Specifically, Answer Set Prolog is used to represent the incomplete domain knowledge, and to reason with this knowledge for planning and diagnostics. Unexpected state transitions during plan execution trigger reinforcement learning to interactively discover specific (i.e., ground) instances of previously unknown axioms. A decision tree induction approach and the relational representation encoded in the Answer Set Prolog program are used to generalize from these discovered axioms, providing generic axioms that revise the existing Answer Set Prolog program and are thus used for subsequent reasoning. The architecture's capabilities are illustrated and evaluated in a simulated domain that has an assistive robot moving particular objects to desired locations or people in an office.
