Learning for Mobile-Robot Error Recovery (Extended Abstract)

Introduction Mobile defense, rescue, and space robots are expected to operate in rough and unknown terrain, typically with the aid of human teleoperation. However, teleoperation alone cannot eliminate robots experiencing “locomotion errors,” such as when a robot’s leg, wheel, or body becoming trapped in a crevice. For example, NASA’s Mars rover Spirit became trapped in soft ground and could not be released even after eight months of teleoperated error recovery attempts (Wolchover 2011). Such locomotion errors occur when the robot’s circumstances differ from the original design conditions. Indeed, this problem can be further generalized—as robots become more common and operate over long periods (Marder-Eppstein et al. 2010), even personal robots such as the Roomba and the PR2 can be rendered immobile in unpredictable home environments. When a locomotion error occurs, human operators are required to assist robots by, (1) diagnosing the specific problem and, (2) finding a escape strategy (i.e., a sequence of actions) to release the robot. Current automatic motion planning and control algorithms cannot handle the complex robot-environment physical interactions and dynamic movements required to find an escape route for stuck robots, such as running into obstacles to change the environment, and instead focus on quasi-static and contact-free planning. Our long-term goal is to significantly improve the robustness of mobile-robots by allowing them to autonomously recover from locomotion errors. A learning approach is ideally suited to this problem since the robots are required to extricate themselves from new scenarios about which little information is available a priori. This paper introduces the novel problem of autonomous mobile-robot error recovery. First, it will define the mobilerobot error recovery problem and its scope. Second, it enumerates the major challenges faced when tackling this problem. Third, it lays out a generic framework for studying the autonomous mobile-robot error recovery problem.