Multiworld Motion Planning

Abstract

While predictive planning approaches have had broad success in robot navigation, most implementations of them have important limitations. In particular, they plan around a single predicted outcome (plus uncertainty) or the union of several likely outcomes. Here, we introduce the multiworld motion planning problem, where a robot plans around each likely outcome separately, while maintaining a safe trajectory until it can determine which of the predictions actually happened. We introduce a method both for split detection (understanding when multiple plans can be created) and split planning (planning path over these separate contingencies). We apply our techniques to simulated robots with both holonomic and nonholonomic dynamics in a variety of scenarios. We show that our resulting approach can produce substantially safer and shorter trajectories than traditional motion planning in critical situations that commonly arise in tasks such as navigating in human environments.