Solving Motion and Action Planning for a Cooperative Agent Problem Using Geometry Friends

摘要

In this paper we discuss the development of agents for the Geometry Friends game, which poses simultaneously problems of planning and motion control in an physics-based puzzle and platform 2D world. The game is used in a competition, held yearly, that challenges participants to solve single player and cooperative levels. Our work addresses the two. The approach followed uses Rapidly-Exploring Random Trees with strategies to accelerate the search. When comparing with other agents on the competition, our results show that our agents can solve the single player challenges without overspecialization and are also promising for the cooperative levels with either agent-agent and human-agent players.