AI Learns to Play Pokémon Go on AI Sandboxes (zozo123.github.io)

A tech enthusiast has successfully developed a novel system that enables large language models (LLMs) to "play" Pokémon Go-style games within isolated sandboxes, despite significant restrictions posed by the game's architecture. Due to Play Integrity hardware verification and server-side bans targeting bot-like behavior, traditional approaches to direct gameplay are unfeasible. Instead, the developer utilized genetic algorithms to evolve the system prompts of agents within forked virtual machines, allowing for a simulated environment to assess and refine gameplay strategies. This unique architectural setup employs Pokémon Crystal as a training ground, utilizing RAM-derived fitness signals from gameplay to guide the evolution of agent behaviors. This breakthrough is significant for the AI/ML community as it highlights innovative methods for parallel processing and evolutionary learning in scenarios where traditional reinforcement learning might lead to quick account bans or security concerns. The use of multi-agent systems within parallelized sandboxes encourages collaboration and competition among evolving agents, leading to improved performance without direct interaction. The system has demonstrated consistent upward trends in agent fitness across generations, successfully enabling the agents to accomplish objectives such as capturing Pokémon and navigating game maps—all while circumventing the game's strict limitations. This approach may inspire future explorations in using genetic algorithms for broader applications within AI.