Distributed Fleet Management in Noisy Environments via Model-Predictive Control

We consider dynamic route planning for a fleet of Au-tonomous Mobile Robots (AMRs) doing fetch and carry taskson a shared factory floor. In this paper, we propose StochasticWork Graphs (SWG) as a formalism for capturing the seman-tics of such distributed and uncertain planning problems. Weencode SWGs in the form of a Euclidean Markov DecisionProcess (EMDP) in the tool UPPAALSTRATEGO, which em-ploys Q-Learning to synthesize near-optimal plans. Further-more, we deploy the tool in an online and distributed fashionto facilitate scalable, rapid replanning. While executing theircurrent plan, each AMR generates a new plan incorporat-ing updated information about the other AMRs positions andplans. We propose a two-layer Model Predictive Controller-structure (waypoint and station planning), each individuallysolved by the Q-learning-based solver. We demonstrate ourapproach using ARGoS3 large-scale robot simulation, wherewe simulate the AMR movement and observe an up to 27.5%improvement in makespan over a greedy approach to plan-ning. To do so, we have implemented the full software stack,translating observations into SWGs and solving those withour proposed method. In addition, we construct a benchmarkplatform for comparing planning techniques on a reasonablyrealistic physical simulation and provide this under the MITopen-source license.