THIS PROPOSAL AIMS TO CODIFY PAST EXPERIENCE IN THE FORM OF A GENERATIVE MODEL OF INTERACTION THAT USES THE MEMORY OF PAST INTERACTIONS TO MAKE STATE PREDICTIONS INTO THE FUTURE. WE ARGUE THAT THIS CONSTITUTES A POWERFUL FORM OF BACKGROUND KNOWLEDGE THAT CAN BE USED TO COMPENSATE FOR MISSING ENVIRONMENTAL STRUCTURE BY: ACTIVELY GATHERING INFORMATION; MONITORING UNFOLDING BEHAVIOR RECOGNIZING AND EXPLOITING STABLE PATTERNS AFFORDED BY THE ENVIRONMENT; AVOIDING ACTIONS (AND STATES) THAT LEAD TO UNRECOVERABLE FUTURE STATES; AND SUPPORTING A LARGE VARIETY OF LEARNING AND PLANNING ALGORITHMS. BACKGROUND KNOWLEDGE IN MEMORY IS USED IN THIS PROPOSAL AS A MEANS OF QUANTIFYING CONFIDENCE AND RISK AND ELECTING CONSERVATIVE INFORMATION GATHERING ACTIONS TO IMPROVE CERTAINTY IN THE SAME DECISION PROCESS THAT CONSIDERS OTHER BEHAVIOR. WITH SUCH BACKGROUND KNOWLEDGE WE PROPOSE THAT ROBOTS CAN ACTIVELY PERCEIVE IMPORTANT ASPECTS OF AUTONOMOUS INTERACTIONS AND THEREBY: CONTROL RISK LEVERAGE PRIOR EXPERIENCE TO SOLVE NOVEL TASKS AND ACHIEVE AN UNPRECEDENTED LEVEL OF LONG-TERM AUTONOMY. WE OUTLINE A RESEARCH PLAN TO PRODUCE AN INTEGRATED COGNITIVE PERCEPTUAL AND MOTOR SYSTEM THAT LEARNS QUICKLY RESPONDS ACTIVELY TO UNCERTAINTY CREATES CONTINGENCIES FOR UNEXPECTED STATE TRANSITIONS AND TRANSFERS WHAT IT KNOWS EFFICIENTLY TO NEW PROBLEMS. A NEW FRAMEWORK FOR AUTONOMOUS CONTROL IS PROPOSED IN WHICH SENSORY AND MOTOR RESOURCES PARAMETERIZE A LANDSCAPE OF ATTRACTORS DESCRIBING AN ``OBJECT." THE RESEARCH PROPOSED WILL STUDY LEARNING ALGORITHMS FOR ACQUIRING AFFORDANCE MODELS DEFINED AS RELIABLE STATE TRANSITIONS AT MULTIPLE LEVELS OF ABSTRACTION. IN THE COURSE OF INTERACTING WITH OPEN ENVIRONMENTS MANY SUCH MODELS COULD EMERGE SO OUR RESEARCH PLAN WILL EXAMINE PRACTICAL DATA-DRIVEN METHODS FOR CONTROLLING THE COMPLEXITY OF LEARNING---WE CONSIDER SEEDING MODELS USING A TYPE OF LEARNING FROM DEMONSTRATION (LFD) AND MAKING THEM MORE COMPREHENSIVE USING AN INTRINSICALLY MOTIVATED STRUCTURE LEARNING (IMSL) METHOD TO LEARN TASK-INDEPENDENT TRANSITION MODELS BETWEEN ATTRACTORS. POPULATIONS OF THESE MODELS ARE USED TO PARSE SENSORY FEEDBACK FROM COMPLICATED AND UNSTRUCTURED ENVIRONMENTS TO DIRECT EXPLORATION WHEN STATE FEEDBACK IS AMBIGUOUS AND EXPLOITATION WHEN OUTCOME STATES ARE LIKELY TO BE SAFE AND PRODUCTIVE. FINALLY WE PROPOSE AN EXPERIMENTAL EVALUATION OF LEARNING AND PLANNING IN THIS REPRESENTATION IN THE CONTEXT OF AUTONOMOUS ROBOT GRASPING AND MANIPULATION WITH A BIMANUAL MOBILE MANIPULATOR THAT INTERACTS EFFECTIVELY WITH BOTH FAMILIAR AND UNFAMILIAR OBJECTS.