Graphs of Convex Sets for Lifelong Manipulation Motion Planning

Abstract

Manipulation has great potential not only for industrial settings with predefined trajectories, but also for tasks such as human-robot interaction, where more sophisticated planning methods are required. One promising direction is planning with Graphs of Convex Sets (GCS) in the configuration space. This approach has been shown to be effective on many robotic manipulation tasks, but it still has a major bottleneck: constructing convex sets in high-dimensional setups is computationally expensive. Moreover, obtaining a sufficiently good decomposition of the whole scene in advance can be difficult and time-consuming. To this end, in this paper we focus on the online planning setup in which goal states are sent to the system iteratively. Instead of requiring a good decomposition of the entire scene beforehand, we refine and expand the graph of convex regions only when needed, as new goal states arrive. In this way, the decomposition is built incrementally rather than fully precomputed. We evaluate our approach on different planning scenes and compare it with sampling-based and optimization-based baselines in terms of planning time and trajectory quality. The results show that our method is faster while remaining competitive in trajectory quality.