THE USE OF NEURAL NETWORKS FOR PROGRAMMING AND CONTROL OF SENSORY BASED ROBOTICS

TWO MAJOR IMPEDIMENTS TO THE EFFECTIVE USE OF ROBOTICS ARE 1) THE DIFFICULTY IN DEVELOPING ROBUST, ADAPTIVE, GOAL-DIRECTED BEHAVIOR BASED ON SENSORY INTERACTIVE CONTROL AND 2) THE DIFFICULTY IN REPROGRAMMING A ROBOTIC SYSTEM TO CARRY OUT NEW TASKS. THE OBJECTIVE OF THIS WORK IS TO DEVELOP A HYBRID CONTROL SYSTEM MERGING TRAINABLE NEURAL NETWORK MODULES WITH A WELL-STRUCTURED KNOWLEDGE BASED CONTROL ARCHITECTURE. THE CONTROL STRUCTURE WILL MAINTAIN TASK DECOMPOSITION CONTROL AND DATA MANAGEMENT WHILE THE NEURAL NET MODULES WILL PROCESS THE SENSOR DATA AND CONTROL THE ACTUATOR SERVOS. RESEARCH AND ANALYSIS WILL BE DONE IN THE AREAS OF 1) CHARACTERIZING THE RELATIONSHIP OF THE ADJUSTABLE NETWORK PARAMETERS TO THE TYPE OF FUNCTION THAT CAN BE STORED, THE SPEED OF TRAINING, TYPE OF TRAINING, ACCURACY OF THE STORED FUNCTION AND TIME SEQUENCING AND TIME DELAY EFFECTS; 2) DEFINING THE BOUNDARIES BETWEEN NEURAL NETWORK PROCESSING AND MORE CONVENTIONAL TASK REPRESENTATION; 3) DEFINING METHODS TO EXTRACT WORLD MODEL DESCRIPTIONS FROM THE TRAINEDNEURAL NETS. DEMONSTRATION TASK EMPLOYING HARDWARE IMPLEMENTATIONS WILL BE USED TO VALIDATE RESULTS. IT IS EXPECTED THAT PHASE I WILL PROVIDE SUFFICIENT CHARACTERIZATION OF THE NEURAL NETWORK AND CONTROL TASKS TO ALLOW PHASE II PROTOTYPE OF A TESTBED HYBRID CONTROL SYSTEM. PHASE III WILL DEVELOP THE TESTBED INTO A GENERIC LOW-COST TRAINABLE PC-BASED TOOL SET TO ALLOW FAST PROGRAMMING SPECIFICATION OF COMPLEX SENSORY INTERACTIVE TASKS.