Rethinking perception-action loops via interactive perception and learned representations

2018-03-27For robots to become fully autonomous in real-world environments, they must be able to cope with uncertainties resulting from imperfect control and perception. Although robotic perception and control systems typically account for uncertainty independently, they do not consider uncertainty in the coupling between perception and control, leading to degraded performance. To address this shortcoming, recent approaches in robotics follow the insight that interaction with the environment can facilitate perception and inform manipulation, which results in reducing the uncertainty between perception and control components. ❧ Another challenge that is at the intersection of these components involves intermediate representations used for the transition between the two. Traditionally, the interfaces between perception and control systems have been well-defined and hand-specified by robotics engineers, which might cause the intermediate representations to be constraining. This is in strong contrast to modern, flexible and learned representations that are optimized for the task at hand using deep learning. ❧ In this thesis, we present several methods that demonstrate how interaction with the environment, closed perception-action loops and learned representations are beneficial for manipulation and perceptual tasks. Specifically, we look at the problems of Interactive Perception, where the robot forcefully interacts with the environment, Active Perception, where the robot changes its sensor states to acquire additional data, and Robot Learning, where the robot learns the intermediate representations in an end-to-end fashion. We show applications of these paradigms to the tasks of articulation model estimation, grasping, multi-robot target tracking, self-calibration, locomotion and acquiring manipulation skills