Learning-based AI pursuit in dynamic environments with environmental constraints

This research explores the efficiency of reinforcement learning (RL) for developing intelligent pursuit AI in dynamic 3D game environments with challenges like slippery surfaces and obstacles. We developed an RL agent trained via curriculum learning to help it gradually learn how to avoid obstacles and take advantage of the environment’s physics. The RL agent shows promising adaptability and unique capabilities, particularly in handling slippery surfaces. Its performance was benchmarked against a tradi- tional NavMesh agent. While the NavMesh agent generally achieved better results in speed and safety, almost always reaching the target in 4.63s, the RL agent, on the other hand, though generally slower (4.71s-5.02s), it showed promising adaptability in different environments. The RL agent’s performance depended a lot on how the environment was set up; removal of a single initial obstacle significantly improved its performance and reduced collisions. This shows RL’s ability to handle complex interactions, but also shows that it might struggle to consistently beat traditional methods.