Hybrid Risk-Aware Conditional Planning with Applications in Autonomous Vehicles

In this paper, we address the problem of risk-aware conditional planning where the goal is generating risk bounded motion policies in the presence of uncertainty. The problem is modeled as a chance-constrained Partially Observable Markov Decision Process (CC-POMDP) with one controllable agent and multiple uncontrollable agents, each of which can choose from a set of maneuver actions. The risk is defined as the probability of the controllable agent violating safety constraints. Off-line computations include generating a library of probabilistic maneuvers for the controllable agent and planning an initial motion policy to execute. During runtime, the conditional planner can quickly look up maneuver sequences to ensure risk bounds as the world around our agent evolves. We introduce the iterative RAO* heuristic search algorithm, which iteratively generates risk bounded conditional plans over a finite horizon. We demonstrate the performance of the provided approach on two planning problems of autonomous vehicles.