Overcooked-AI: Multi-Agent Reinforcement Learning Project This project demonstrates a full implementation of Multi-Agent Reinforcement Learning (MARL) in the Overcooked-AI environment, a widely used benchmark for cooperative AI research. Project Summary • Implemented a MAPPO-based (Multi-Agent PPO) learning pipeline in Python. • Built a custom Gym wrapper for Overcooked-AI (2021 stable version). • Designed reward shaping, feature extraction, trajectory collection, and PPO training loops. • Agents successfully learned coordinated behaviors such as ingredient delivery and soup production. Technical Highlights • Environments: Overcooked-AI (Cramped Room, Asymmetric Advantages). • Algorithms: PPO / MAPPO, baseline rule-based agents for comparison. • Tools: PyTorch, NumPy, custom policy networks, action masking, logging utilities. • Features: Multi-agent state featurization, advantage estimation, value normalization. Results The learning curve below represents an example of agent improvement over early episodes. Agents progressively reduce collisions, optimize movement, and increase team reward.
- AI Engineer