Session
Weekend Session VIII: Advanced Technologies - Research & Academia 2
Location
Utah State University, Logan, UT
Abstract
Conventional Earth-observing satellites spend most of their time taking cloudy imagery as two-thirds of Earth’s surface is covered by clouds at any one time. To combat this, a mission concept called dynamic tasking has been proposed. In this mission, a low-resolution, wide field-of-view lookahead scans upcoming tasks. Modern computer vision and onboard compute can then rank these upcoming tasks, which can be used to dynamically re-optimize the schedule onboard. The vision system can additionally be used to add tasks into the schedule for low-frequency phenomena. While previous studies have focused on steerable lookahead instruments (e.g. radar) that can effectively be decoupled from body pointing, this work examines the use of a body-fixed lookahead sensor such as a conventional camera. The coupling of the lookahead sensor’s imaging space with the spacecraft’s attitude and the necessary settling time before primary sensor operation presents additional complications in scheduling. This work looks at using reinforcement learning to develop lookahead heuristics, as an input to a classical scheduler. The development of effective lookahead policies could enhance the capability of smaller spacecraft by facilitating low-cost dynamic tasking, thereby increasing the throughput of valuable data and bridging the gap between monitoring and tasked satellite missions.
Reinforcement-Learned Lookahead Heuristics for Earth-Observing Satellites
Utah State University, Logan, UT
Conventional Earth-observing satellites spend most of their time taking cloudy imagery as two-thirds of Earth’s surface is covered by clouds at any one time. To combat this, a mission concept called dynamic tasking has been proposed. In this mission, a low-resolution, wide field-of-view lookahead scans upcoming tasks. Modern computer vision and onboard compute can then rank these upcoming tasks, which can be used to dynamically re-optimize the schedule onboard. The vision system can additionally be used to add tasks into the schedule for low-frequency phenomena. While previous studies have focused on steerable lookahead instruments (e.g. radar) that can effectively be decoupled from body pointing, this work examines the use of a body-fixed lookahead sensor such as a conventional camera. The coupling of the lookahead sensor’s imaging space with the spacecraft’s attitude and the necessary settling time before primary sensor operation presents additional complications in scheduling. This work looks at using reinforcement learning to develop lookahead heuristics, as an input to a classical scheduler. The development of effective lookahead policies could enhance the capability of smaller spacecraft by facilitating low-cost dynamic tasking, thereby increasing the throughput of valuable data and bridging the gap between monitoring and tasked satellite missions.
