Planning under Uncertainty from Behaviour Trees

Behaviour trees (BTs) are popular within robotics due to their reactivity, reusability, and modularity. BTs are often designed by hand using expert domain knowledge. However, robot environments contain sources of uncertainty which affect robot behaviour. It is challenging for human designers to reason over the effects of uncertainty up to the task horizon, limiting robot performance. For example, the chance of an unexpected blockage late along a robot’s route should encourage the robot to take an alternate path. Therefore, in this paper we refine the task-level behaviour encoded in a BT through planning under uncertainty. The refinement process modifies when action nodes are executed by reasoning over the effects of uncertainty, improving task performance. We first extract a state space from the BT and learn a set of Bayesian networks (BNs) which model the stochastic dynamics of robot actions. We then use the extracted state space and BNs to construct and solve a Markov decision process which captures robot execution. This produces a policy which describes the refined behaviour. We empirically demonstrate how our approach reduces the completion time for robot navigation and search tasks.