Robust Control for Robots via Minimal-Information Policies

This study investigates a fundamental trade-off between a robot's state information usage and optimal decision making. Typically, robot state estimators and control systems are designed to produce and consume as much state information as possible — yielding a tight coupling between sensing and control.

This work identifies that it is often advantageous for the robot to minimize information usage to limit the effect of sensing error on control performance.We formalize the robot's control objective as a rate-distortion problem, and demonstrate that the optimal controller is given by a Gibbs measure. This structure allows us to quantify the performance of the controller when composed with a bounded-error state estimator.

Finally, we optimize the robot's information usage for a fixed level of estimation error and see that higher error levels produce lower information controllers.We conclude the study with corroborating simulations. Scenarios include hopping robots, ball catching, and object manipulation. Each example, demonstrates that low-information controllers perform robustly under estimation uncertainty.