Adaptive Introspection and Deployment for Robust Long Duration Autonomy
Nathan Michael, Sebastian Scherer · 2014
Long duration autonomy for unmanned systems is difficult to achieve as current systems are limited to anticipated exceptions and do not adapt to long-term changes in the environment.The project goal is to enable long-term operation in unpredictable environments through adaptive introspection and deployment approaches, emphasizing vehicle and system level adaptation and robustness as a robotic team cooperates toward a common and persistent mission goal.At the vehicle level, we address the problems of identifying unexpected vehicle states and developing robust mitigation policies and behaviors through introspection.At the system level, we propose long-term planning methodologies that coordinate the robot team toward a common mission objective while learning, adapting to, and anticipating changing vehicle and environment conditions.The proposed research outcomes will be broadly applicable to unmanned vehicle systems operating over long time horizons to achieve a common and persistent mission.In order to evaluate the performance of the proposed methods, we also pursue an integration and experimentation design that will provoke exceptions over long durations. OBJECTIVESThe project is organized into three research tasks.The first two tasks pursue vehicle and system level adaptation, respectively.The third task develops an integrative experimental framework toward evaluating the approaches developed through the first two tasks.Task 1: Introspection for Robust Autonomy: Typical robot systems are only robust to known contingencies, and therefore cannot adapt to unknown contingencies.Our goal is to be able to react to unknown exceptions through adaptation in the configuration space of sensors, processes, and actuators.Processes represent sets of viable algorithms that have varying applicability, such as a set of visual odometry algorithms for monocular and stereo situations, that will be automatically selected through introspection to enable robust autonomy and reaction to unknown contingencies.We consider two research directions to achieve this goal:• Introspection of the effectiveness of sub-components and their parameters towards mission progress, and • Configuration changes based on the effectiveness of sub-components of the system.