An Online POMDP Solver for Uncertainty Planning in Dynamic Environment

Abstract

Motion planning under uncertainty is important for reliable robot operations in uncertain and dynamic environments. Partially Observable Markov Decision Process (POMDP) is a general and systematic framework for motion planning under uncertainty. To cope with dynamic environment well, we often need to modify the POMDP model during runtime. However, despite recent tremendous advances in POMDP planning, most solvers are not fast enough to generate a good solution when the POMDP model changes during runtime. Recent progress in online POMDP solvers have shown promising results. However, most online solvers are based on replanning, which recompute a solution from scratch at each step, discarding any solution that has been computed so far, and hence wasting valuable computational resources. In this paper, we propose a new online POMDP solver, called Adaptive Belief Tree (ABT), that can reuse and improve existing solution, and update the solution as needed whenever the POMDP model changes. Given enough time, ABT converges to the optimal solution of the current POMDP model in probability. Preliminary results on three distinct robotics tasks in dynamic environments are promising. In all test scenarios, ABT generates similar or better solutions faster than the fastest online POMDP solver today; using an average of less than 50 ms of computation time per step.