Multi-stage Perception-aware Chance-constrained MPC with Applications to Automated Driving

Abstract

Perception-aware Chance-constrained Model Predictive Control (PAC-MPC) accounts for the interdependence between perception and control for systems operating in uncertain environments. The environment is discovered by perception, which imposes chance constraints on system operation. PAC-MPC can handle a perception quality that depends on the system states and/or inputs, thus affecting uncertainty quantification in the chance constraints. In this paper, we extend PAC-MPC by introducing a scenario-based prediction for future measurements, so that the resulting multi-stage PAC-MPC does not require a conservative estimate of the measurement prediction error covariance. We demonstrate PAC-MPC for automated vehicle control when obstacles and road boundaries are uncertain and perceived by variable precision sensors subject to an overall sensing budget and when the scenarios are generated based on possible obstacle behaviors.