PAMPR: Predictability-Aware Motion Planning for Robots

Motion planning algorithms are crucial to the effective operation of robots; however, many motion planning algorithms exist, and they have complementary strengths when applied to different planning scenarios. The goal of this work is to increase predictability in the context of motion planning across a set of targets. A challenge is that each task and each planner have different levels of variability in the time they need for each target. The sub-goals of this work are therefore to shorten the time to complete the sequence at the same time as reducing the variability of the completion time. This combination helps reduce the likelihood that a deadline is missed. Our framework for predictability-aware motion planning for robots (PAMPR) uses machine learning to predict the planning and execution time for each available algorithm and query; these predictions are then used by a constraint solver to find an optimal sequence of targets and algorithm choices. We use an assistive robot arm example to show that this approach can lead to faster and more predictable completion times for the collection of tasks. Compared to the single best planner according to the training data, PAMPR reduces the right tail of the distribution by 47% and is twice as fast on average.