THE GOAL OF THIS PROJECT IS TO DEVELOP AND DEMONSTRATE A RISK-SENSITIVE FRAMEWORK FOR LEARNING AND DECISION-MAKING THAT ALLOWS INTELLIGENT PHYSICAL SYSTEMS (IPS) TO OPERATE IN UNKNOWN AND UNCERTAIN ENVIRONMENTS WITHOUT CONTINUOUS CONTROL OR SUPERVISION. TYPICALLY THE GROUND CONTROLLERS OF SPACE ASSETS CHOOSE ACTIONS THAT ARE EITHER "EXPLOITATIVE " THAT IS THEY SEEK TO EXPLOIT A MODEL OF THE ENVIRONMENT OR SYSTEM STATE TO FULFILL A GIVEN TASK OR ARE "EXPLORATORY " THAT IS THEY SEEK TO REDUCE MODEL UNCERTAINTY VIA EXPLORATION E.G. OF NEW STATES. THIS PROPOSAL SEEKS TO DEVISE A NOVEL ALGORITHMIC FRAMEWORK THAT ALLOWS AN IPS TO OPTIMALLY AND AUTONOMOUSLY BALANCE EXPLOITATION AND EXPLORATION ACTIONS WHILE ALSO ACCOUNTING FOR RISK. THE SPECIFIC TECHNICAL OBJECTIVES ARE: -- OBJECTIVE 1. PROBABILISTIC DYNAMICS MODELS: BY ADOPTING A BAYESIAN APPROACH TO ONLINE LEARNING DESIGN A PROBABILISTIC MODELING FRAMEWORK THAT CAN EXPRESS PRIOR UNCERTAINTY OVER MODELS OF THE ENVIRONMENT AND SYSTEM DYNAMICS AND CAN RAPIDLY ADAPT TO INCORPORATE DATA GATHERED ONLINE. -- OBJECTIVE 2. LEARNING AND DECISION-MAKING ALGORITHMS: DEVISE LEARNING AND DECISION-MAKING ALGORITHMS THAT LEVERAGE SUCH A PROBABILISTIC REPRESENTATION OF THE ENVIRONMENT AND SYSTEM DYNAMICS TO GENERATE RISK-SENSITIVE CONTROL ACTIONS IN A SCALABLE MANNER. -- OBJECTIVE 3. VALIDATION OF FRAMEWORK ON KEY IPS SCENARIOS: VALIDATE MODELS AND ALGORITHMS ON TWO REPRESENTATIVE SCENARIOS OF INTEREST TO NASA SPECIFICALLY (1) AUTONOMOUS TRAVERSABILITY ASSESSMENTS FOR ROVER NAVIGATION AND (2) ROBUST GRASPING AND MANIPULATION OF NON-COOPERATIVE FREE-FLOATING OBJECTS. THE KEY INNOVATION OF THIS PROJECT IS TO LEVERAGE RECENT ADVANCES IN MACHINE LEARNING FOR DECISION-MAKING AND CONTROL FROM THE ROBOTICS COMMUNITY TO THE PLANNING AND CONTROL OF AUTONOMOUS SPACECRAFT AND SPACE ROBOTS. SPECIFICALLY BY ADOPTING A BAYESIAN APPROACH THIS PROJECT WILL GENERATE NEW TOOLS FOR PROBABILISTIC DYNAMICS MODELING ROOTED IN DEEP LEARNING AND META LEARNING AS WELL AS NEW METHODS FOR RISK-SENSITIVE DECISION-MAKING COMBINING BAYES-ADAPTIVE PLANNING RISK THEORY COMPUTATIONAL GAME THEORY AND REINFORCEMENT LEARNING. THIS PROPOSAL BRINGS TOGETHER A TEAM OF EXPERTS IN DECISION-MAKING UNDER UNCERTAINTY MACHINE LEARNING FOR CONTROL OPTIMAL CONTROL THEORY AND SPACE ROBOTICS FROM STANFORD AND UC BERKELEY. IT BUILDS ON RECENT RESULTS BY THE INVESTIGATORS WHICH COLLECTIVELY PROVIDE MODELS AND ALGORITHMS TO ALLOW ROBOTIC SYSTEMS TO QUICKLY AND EFFECTIVELY ADAPT ONLINE TO NEW SITUATIONS. ADDITIONALLY IT BUILDS ON SEVERAL ONGOING COLLABORATIONS WITH NASA WHICH WILL BE INSTRUMENTAL TO A POSSIBLE TECHNOLOGY INFUSION IN NASA MISSIONS. THE INVESTIGATORS WILL BE ASSISTED BY ONE FULL-TIME AND THREE PART-TIME GRADUATE STUDENTS FROM STANFORD AND UC BERKELEY. THIS PROPOSAL RESPONDS TO TOPIC 2: "SMART AND AUTONOMOUS SYSTEMS FOR SPACE." IT DIRECTLY ADDRESSES THE INTEREST FROM NASA IN DEPLOYING SYSTEMS THAT CAN OPERATE AUTONOMOUSLY IN THE UNCERTAIN ENVIRONMENT OF SPACE WITHOUT THE NEED FOR CONTINUOUS SUPERVISION BY GROUND CONTROL. WE WILL APPLY OUR ALGORITHMIC FRAMEWORK TO TWO SPACE APPLICATIONS FOR WHICH THE PI HAS EXTENSIVE EXPERIENCE AND WHICH ARE HIGHLIGHTED IN THE 2015 NASA TECHNOLOGY ROADMAPS REPORT: (1) AUTONOMOUS TRAVERSABILITY ASSESSMENTS FOR ROVER NAVIGATION CORRESPONDING TO SURFACE MOBILITY #4.2.5 AND SMALL-BODY AND MICROGRAVITY MOBILITY #4.2.4 (A LONG-STANDING COLLABORATION WITH NASA JPL FIRST INITIATED AS A NASA NIAC PROJECT) AND (2) AUTONOMOUS GRASPING AND MANIPULATION OF NON-COOPERATIVE FREEFLOATING OBJECTS CORRESPONDING TO GRAPPLING #4.3.7 (THE SUBJECT OF AN ONGOING NASA ESI AWARD).