AI Learns to Escape Extreme Maze

An AI agent named Albert learns to navigate increasingly complex mazes through reinforcement learning and trial-and-error. Starting with random movements, Albert gradually improves through 221,700+ attempts across multiple maze types including trap mazes, lava mazes with jumping mechanics, and moving obstacle courses with wind and spinning hazards. The training demonstrates key AI learning concepts like sparse rewards, learned behaviors that are difficult to unlearn, and the challenge of complex sequential decision-making in game environments.