Rapidly-Exploring Random Tree Approach for Geometry Friends

The paper presents a two-stage agent design for Geometry Friends, a physics-based cooperative platform game, using a Rapidly-exploring Random Tree (RRT) for high-level planning and a PID controller for real-time execution. Its central argument is that prior winning approaches were overspecialized to public levels or depended on extensive learning, whereas a generic, map-feature-independent method should generalize to unseen private levels. The authors frame their contribution around this generality, positioning the planner and controller as reusable components rather than tailored heuristics.

The evaluation centers on the 2015 IEEE Computational Intelligence and Games competition. The circle agent placed first on the circle track with a score of 2096, while the rectangle agent placed third on the rectangle track with 1487. Success rates were 0.32 on public circle levels and 0.38 on private levels, a comparison the authors cite as evidence that performance does not collapse on unseen maps. Shortfalls were attributed primarily to planning failures involving non-adjacent platforms at differing heights, along with some control inaccuracies. A dedicated release routine handles stuck states, and special actions such as jumping, tilting, and morphing are managed explicitly.

The authors conclude that their approach was competitive against more advanced algorithms while remaining general, and they argue that the agents could in principle solve any level. They acknowledge remaining weaknesses and plan an improved submission for the 2016 competition, with an extension to the cooperative track. Beyond the game, they note that the framework adapts to other problems by replacing problem-specific functions within the RRT, and that the PID controller can be reused for point-to-point movement along a single axis.