ISAS Space Science Colloquium & Space Science Seminar
Deep and Shallow Learning Models for Autonomous Space Exploration
Roberto Furfaro（ロベルト フルファロ）
University of Arizona
Autonomous and unconstrained exploration of small and large bodies of the solar system requires the development of a new class of intelligent systems capable of integrating in real-time stream of sensor data and autonomously take optimal decisions, i.e. decide the best course of action. For example, future missions to asteroids and comets will require that the spacecraft be able to autonomously navigate in uncertain dynamical environments by executing a precise sequence of maneuvers (e.g. hovering, landing, touch-and-go) based on processed information collected during the close-proximity operations phase. Currently, optimal trajectories are determined by solving optimal guidance problems for a variety of scenarios, generally yielding open-loop trajectories that must be tracked by the guidance system. Although deeply rooted in the powerful tools from optimal control theory, such trajectories are computationally expensive and must be determined off-line, thus hindering the ability to optimally adapt and respond in real-time to 1) uncertainties in the unknown dynamical environment; 2) detected hazards; and 3) science value analysis.
Over the past few years, there has been an explosion of machine learning techniques involving the use of shallow and deep neural networks to solve a variety of problems spanning from object detection to image recognition to natural language processing and sentiment analysis. The recent success of deep learning is due to concurrent advancement of the fundamental understanding on how to train deep architecture, the availability of large amount of data and critical advancements in computing power (e.g. extensive use of GPUs). One can naturally ask the following: how can such techniques help the development of the next generation of robust and adaptive algorithms that may enable autonomous space exploration? In this talk, I will address this problem by presenting a variety of methods and techniques that have been recently developed by my research team in the context of autonomous planetary landing and close proximity operations around small bodies. The methodologies span from supervised learning to deep reinforcement learning and demonstrate that such approaches may be implemented to enable intelligent autonomous systems for both guidance, control and real-time decision-making during the robotic exploration of the solar system."
Place: Bldg. A 1F Nyusatsu Conf. room（1134）/ A棟１F入札・会議室（1134号室）