Automatic detection of abnormal motion

Abstract

Modern robots are almost helpless in the real world despite their unprecedented perception. They struggle to react appropriately to random events. They require human supervisors to detect and react to situations for which they are not explicitly prepared. To survive in the real world without human supervision, robots must autonomously distinguish normal and abnormal situations. The abstract nature of this task has no simple qualitative solution, which prohibits an algorithmic solution and mandates human intervention. This thesis presents a method for robots to autonomously identify abnormal motions. The method requires a human operator to guide a robot through the motions it is expected to perform during normal operation, which allows the robot to group recurring motions with a clustering algorithm. The resulting clusters form a quantitative model of “normal motion”. This model allows the robot to calculate the probability that an observed motion is a normal motion. Results demonstrated that abnormal motions such as being lifted, pushed or stepped on, were correctly indicated with low probabilities, and that motions experienced during normal operation were classified as highly probable. This thesis also describes an independent system for robot motion observation. This independent system allows rapid implementation of motion-analysis algorithms on any robot.