Perception-aware model predictive control for constrained control in unknown environments

Abstract

The operation of autonomous systems is inherently constrained by their surrounding environment, which is often time-varying and unknown a priori, necessitating perception using sensors. Hence, control strategies for autonomous systems must take into account the uncertainty of the perceived environment in making decisions, while information acquired by sensors often depends on how the system is operated, e.g., where the sensors are pointed at, or what and how much sensor information is processed. We introduce a perception-aware chance-constrained model predictive control (PAC-MPC) strategy that accounts for the uncertainty of the perceived environment, as well as the dependence of the perception quality on the control actions. The system and the environment are coupled by chance constraints due to the uncertainty in the environment estimate, which depends on control actions. We establish the constraint satisfaction and stability properties of PAC-MPC through appropriate design of the cost function and terminal set, and propose a constructive design procedure for the case of linear dynamics.