AND/OR search techniques for chance constrained motion primitive path planning

Motion primitive planning under parametric uncertainty may be modeled as a chance-constrained Markov Decision Process (CCMDP). Single-query solutions to CCMDPs can be obtained by searching the And/Or graph representing the state-action space of the system. The Risk-bounded AO* (RAO*) algorithm has been proposed as a solution method for this problem, but it scales poorly to MDPs resulting from a motion primitive discretization because it has no mechanism to prioritize expansion of AND nodes. This paper describes an induced heuristic for state-action pairs that can be rapidly computed by leveraging the properties of motion primitives; its value can be used to prioritize AND nodes for more efficient search. Search is further accelerated by leveraging shared symmetry in constraints and dynamics to move almost all computation necessary to enforce convex polytope constraints offline. The performance improvements are demonstrated with path planning problems involving a Dubins Car and a nonlinear aircraft model. (c) 2021 Elsevier B.V. All rights reserved.