Bayesian Active Sensing for Fault Estimation with Belief Space Tree Search

Abstract

Autonomous spacecraft missions must be robust to system component faults such as sensor and actuator failures. An important setting to study fault tolerance is the Bayesian Active Sensing problem, where the system plans control inputs to gain information and estimate system failures quickly and with high confidence. We model the problem as a belief-state planning problem enabling simultaneous estimation of sensor and actuator failures in the presence of noise. Current belief-state tree search planners provide anytime, approximate solutions, but their underlying particle-filter belief update inhibits performance for information-gathering tasks. To address this issue, we propose POMCPMF, a belief-state tree search that uses an exact belief update in the tree search by exploiting the active sensing problem structure to decouple the belief update as a Kalman filter on the physical state and a particle filter on the failure modes. We validate our method on numerical experiments of spacecraft models with unknown sensor and actuator faults to demonstrate (i) the need for an active and planned sensing solution (as opposed to a passive and greedy solution) and (ii) the superior scalability of our method compared to existing active and planned methods. We then demonstrate the applicability of our algorithm to real systems by extending to a non-linear model and deploying our algorithm on a spacecraft simulator robot.

Additional details