Characterizing the Epistemic Uncertainty of Predictive Action Models and Sampling-Based Motion Planners for Robotic Manipulation

Abstract

We derive methods to represent the epistemic uncertainty of models used in long-horizon robot planning problems in autonomous manipulation. We develop a representation of epistemic uncertainty for two types of models: uncertainty over the physical parameters of a model that predicts the observed outcome of a manipulation action and uncertainty over a geometric graph built by a sampling-based motion planner as a representation of the configuration space to answer a motion planning query. We propose a simple planning system that integrates these uncertainty characterizations to reason about the informational value of executing a manipulation action or allocating a number of samples to a sampling-based motion planner.