Decision-theoretic Monte Carlo smoothing for scaling tracking in hybrid dynamic systems

Detecting faults on-board planetary rovers is important since human intervention may not be possible due to communication delays. In This work we propose a scalable method for on-board fault detection and identification that may be applied to general fault models with limited computation. Although our application focus is on diagnosing rover faults, this method is applicable in general for tracking any general non-linear, non-Gaussian hybrid (discrete-continuous) dynamic system online. Our formulation of the fault detection problem requires estimating robot and environmental state, as it changes over time, from a sequence of noisy sensor measurements. We propose a Monte Carlo algorithm that generates new trajectories if the probability of the current set of fault hypothesis being tracked is low. This approach maintains a fixed lag history of measurements, controls and samples. Experimental results of a dynamic simulation of a six-wheel rocker-bogie rover show a significant improvement in performance over the classical approach.