ARCHITECTURES FOR SEMI-AUTONOMOUS PLANNING IN UNRESTRICTED ENVIRONMENTS

THE FOLLOWING PROPOSAL IS FOR EXPLORATORY RESEARCH INTO A PLANNING APPROACH, CALLED BEHAVIORAL PLANNING, WHICH EMPHASIZES ROBUST AGENT PERFORMANCE IN UNRESTRICTED ENVIRONMENTS, WHILE POTENTIALLY SACRIFICING GUARANTEED ABILITY TO ACCOMPLISH GOALS. THIS WORK IS MEANT AS AN ALTERNATIVE TO PLANNERS OF THE LOGIC-INTENSIVE VARIETY WHICHPRODUCE PROVABLY CORRECT ACTION SEQUENCES, BUT HAVE BEEN DEVELOPED EXCLUSIVELY IN THE CONTEXT OF TOY PROBLEMS AND COMPLETELY PREDICTABLE APPLICATION ENVIRONMENTS (WHERE THE AGENT IS FUNCTIONALLY OMNISCIENT). OUR APPROACH OPERATES BYCONSTRUCTING PROGRAMS THAT GOVERN ROBOT ACTION, USING A VOCABULARY OF PRIMITIVE PARTS, CALLED BEHAVIORS, WHICH PERFORM TASKS AND RESPOND TO ENVIRONMENTAL CUES, AND ARE THEN COMBINED IN RESPONSE TO SYSTEM GOALS. THE PRESENCE OF THIS LIBRARY OF PRIMITIVE PARTS IN TURN SUPPORTS INSTRUCTABILITY, THAT IS, THE COMPETENCE TO DESIGN AND MODIFY EXISTING ROBOT PLANS. SINCE NASA APPLICATIONS FOR AIPLANNING TECHNOLOGY ALL, AT TIMES, INVOLVE SUBSTANTIALLY UNPREDICTABLE OR DYNAMIC ENVIRONMENTS, WORK IN THIS GENRE ISALMOST A PREREQUISITE FOR SUPPORTING SPACE STATION AUTOMATION NEEDS. WE PROPOSE A STUDY TO DESIGN AND DEVELOP ONE SUCH PLANNING ARCHITECTURE.