Planning with learned model preconditions for water manipulation

Abstract

Although models are useful for long-horizon reasoning, the complex dynamics of deformable objects in many real-world tasks limit the applicability of many model-based planning algorithms. In this work, we address the difficulty of modeling deformable object dynamics by learning where a set of given high-level dynamics models are accurate: a model precondition. Model preconditions are then used to find trajectories using states and closed-loop actions where the dynamics models are accurate. We demonstrate the efficacy of our approach on pouring and water transport using both analytical and learned models of water movement. We also show qualitative results to provide intuition for what model preconditions might be for the models used in those tasks.