🍲 Continuous OVercooked Simulator in Multi Agent Reinforcemnet Learning

This project implements a multi-agent reinforcement learning environment inspired by Overcooked.
Two agents must cooperate implicitly (without pre-defined roles, they will chose their role by their own decesion based on environment feedback) to complete cooking tasks such as:

• Picking up onions and dishes
• Cooking in a pot
• Delivering finished soup to a counter

📦 Environment Setup

This project uses Unity ML-Agents package with conda envionment package management tool.

1. Create and activate a conda environment (About the python version, you can take reference from Unity Ml-Agents guideline. It's better to set up the environment step by step followed by Unity Ml-Agnets guide)

import COVS.unitypackage to your unity 3D project.

And then create you own environment by conda.

conda create -n overcooked
conda activate overcooked

---

🚀 Training

Run training with your configuration file (run.yaml):

mlagents-learn run.yaml --run-id=<your_run_id> --force

• run.yaml → defines PPO/MAPPO training parameters
• --run-id → unique identifier for each experiment
• --force → overwrite previous runs with the same ID

---

📊 Monitoring with TensorBoard

Start TensorBoard to visualize training curves:

tensorboard --logdir results/<your_run_id>

You can monitor:

• Cumulative Reward → overall performance of agents
• Episode Length → efficiency of task completion
• Policy & Value Losses → training stability

---

🤝 Cooperation

• Two agents interact in the same environment
• Role allocation is implicit, not pre-defined
• Agents must coordinate to achieve maximum reward by finishing and delivering cooked food

---

---

📜 Usage Recap

conda activate overcooked
mlagents-learn run.yaml --run-id=test_run --force
tensorboard --logdir results/test_run