Information-Theoretic Necessary and Sufficient Conditions for the Task-Driven Control of Robots

Abstract

The development of modern robotic systems has significantly benefited from the availability of high-fidelity sensors and efficient computational resources. Together, advances in these areas have led to a significant commercial interest in the applications of autonomous robots for complex tasks. However, while powerful, the high-dimensional sensors now common in robotics (e.g., cameras and LIDAR) pose several challenges for which traditional, general-purpose control and estimation are not well-suited. Specifically, these sensors lack tractable analytical models, and their high-dimensional output spaces are challenging to explore fully empirically or in simulation. Moreover, traditional approaches often result in an unnecessarily tight coupling between the sensing, estimation, and control components of the robot's feedback loop, causing performance degradation when deploying the robot in a new environment due to a shift in the distribution of sensor outputs --- even if the shift only modifies the outputs in a manner that is irrelevant to the robot's task. This dissertation addresses traditional control methods' limitations by establishing task-driven necessary and sufficient conditions for performant feedback control. Necessary conditions quantify how much information a sensor must provide the robot to achieve a performance criterion independent of the controller employed by the robot --- thereby providing a principled method to determine the necessary sensing capabilities of the robot for the task. Sufficient conditions guarantee that a robot will achieve a specific level of performance in new environments by limiting the controller to only depend on \emph{task-relevant} information. Newly developed algorithms for their implementation accompany these conditions. Also included are demonstrations of the efficacy of these methods on problems on common robotics problems featuring high-dimensional sensors.