Hey, gamers and AI enthusiasts! I just came across an amazing project where an AI agent was trained to play the classic arcade game Metal Slug using deep reinforcement learning. The agent, built with Stable-Baselines3 (PPO) and Stable-Retro, receives pixel-based observations and was trained specifically on Mission 1.
What’s fascinating is that the agent faced a tough challenge: dodging missiles from a non-boss helicopter. Despite not being a boss, this enemy became a consistent bottleneck during training due to the agent’s tendency to stay directly under it without learning to evade the projectiles effectively.
After many episodes, the agent started to show decent policy learning — especially in prioritizing movement and avoiding close-range enemies. The goal was to explore how well PPO handles sparse and delayed rewards in a fast-paced, chaotic environment with hard-to-learn survival strategies.
The project’s creator shared a video showcasing the agent’s progress, including a generalization test on Mission 2. It’s impressive to see how the agent adapts to new situations. This project has sparked interesting discussions on training stability, reward shaping, and curriculum learning in retro games.
What do you think about the potential of AI in gaming? Could we see more AI-powered gamers in the future?