RNNs Based Planning with Discrete and Continuous Actions

Anonymous

Keywords: non-convex domains, domains mixed with discrete and continuous actions

Abstract: Dealing with discrete and continuous changes in real-world dynamic environments is of great importance for robots. Despite the success of previous approaches, they impose severe restrictions, such as convex quadratic constraints on state variables, which limits the expressivity of the problem, especially when the problem is non-convex. In this paper, we propose a novel algorithm framework based on recurrent neural networks. We cast the mixed planning with discrete and continuous actions in non-convex domains as a gradient descent search problem. In the experiment, we exhibit that our algorithm framework is both effective and efficient, especially when solving non-convex planning problems.

TL;DR: We propose a novel algorithm based on recurrent neural networks and heuristic searching to handle non-convex domains mixed with discrete and continuous actions.