ML Agents

My version expands the official 2v2 Soccer environment with larger team sizes, updated visuals, and a working scoreboard. The agents learn to play through reinforcement learning. All behaviors are learned, not manually programmed.

I trained the Blue team with an aggressive reward strategy focused on ball chasing and offensive behavior, while the Red team remains unchanged. The model was trained using reinforcement learning in Unity ML-Agents (POCA) for approximately 1.96 million steps. The reward function emphasized forward movement, ball pursuit, and facing the goal.

Compared to the baseline above, the Blue agents in this version demonstrate noticeably more assertiveness and forward pressure.