Overcooked Environment

[mmbajo]

This PR introduces a Overcooked cooking game environment for PufferLib. Sprites were from OvercookedAI

[mmbajo]

Currently training model for demo. Model is not producing meaningful actions. Still debugging.

[mmbajo]

Added intermediate reward shaping to the Overcooked environment to encourage cooperative cooking behavior and provide more frequent learning signals.

Changes

• Onion to pot: +0.1 reward when an agent adds an onion to a pot
• Correct recipe start: +0.1 reward when starting to cook a pot with exactly 3 onions (the target recipe)
• Soup plating: +0.1 reward when transferring a cooked soup from pot to plate
• Dish serving:
  • Correct recipe (3 onions): +5.0 to serving agent, +20.0 to all agents
  • Incorrect recipe: +0.1 to all agents (small consolation reward)

I am now getting decent performance trajectories when training. But I do need some help. Still not sure if this will workout. 🙇

Explained Variance in the positive region! I assume this is a good sign?

Any advice what to try? or change?

[mmbajo]

Hmm... tried training with 1 agent, net can't fully learn how to cook.

[Hadrien-Cr]

do you mind describing your reward structure and rules ? maybe i can help

[mmbajo]

@Hadrien-Cr Hello! I have written a concise README describing the rewards and observations. https://github.com/mmbajo/PufferLib/tree/roze-overcooked-dev/pufferlib/ocean/overcooked