Overcooked
Overcooked-AI is a fully observable cooperative environment where two cooks (agents) must cooperate to prepare and serve onion soups. It is inspired by the popular videogame Overcooked.
We implement all of the original Overcooked-AI layouts: * Cramped Room * Asymmetric Advantages * Coordination Ring * Forced Coordination * Counter Circuit
We also provide a simple method for creating new layouts:
The implementation aims to be as close as possible to the original Overcooked-AI environment, including dynamics, collision logic, and action and observation spaces.
A note on dynamics
In the original Overcooked-AI environment and in this JAX implementation, the pot starts cooking as soon as 3 onions are placed in the pot. An update to Overcooked-AI has since changed the dynamics to require an additional pot interaction to start cooking. Updating the Overcooked-JAX to implement the new pot dynamics is on the roadmap and should be done by the end of 2023.
Action Space
There are 6 possible actions, comprised of 4 movement actions (right, down, left, up), interact and no-op.
Observation Space
The observations follow the featurization in the original Overcooked-AI environment, and is meant to be passed to a ConvNet.
Each observation is a sparse, (mostly) binary encoding of size layout_height x layout_width x n_channels, where n_channels = 26.
For a detailed description of each channel, refer to the get_obs(...) method in overcooked.py.
Reward
JaxMARL's Overcooked reward is the same as the original environment, which corresponds to the score of the game. Specifically, a +1 reward is given to all agents when a recipe is correctly completed and delivered.
Additionally, we include a shaped reward as per the original Overcooked environment. The shaped reward is as follows:
BASE_REW_SHAPING_PARAMS = {
"PLACEMENT_IN_POT_REW": 3, # reward for putting ingredients
"PLATE_PICKUP_REWARD": 3, # reward for picking up a plate
"SOUP_PICKUP_REWARD": 5, # reward for picking up a ready soup
"DISH_DISP_DISTANCE_REW": 0,
"POT_DISTANCE_REW": 0,
"SOUP_DISTANCE_REW": 0,
}
The shaped reward is accessible in the infos returned by the step function of the environment.
Get started
We provide an introduction on how to initialize, visualize and unroll a policy in the environment in ../../tutorials/overcooked_introduction.py.
You can also try the environment yourself by running python interactive.py. Use the arrows to move both agents and the spacebar to interact.
Visualization
We animate a collected set of state sequences.
from jaxmarl.viz.overcooked_visualizer import OvercookedVisualizer
state_seq = [state_t0, state_t1, ...] # collected state sequences
viz = OvercookedVisualizer()
viz.animate(state_seq, env.agent_view_size, filename='animation.gif')
Limitations
Overcooked is an approachable and popular environment to study coordination, but has limitations, notably due to being fully observable. For analysis on this topic, read more here.
Citation
The environment was orginally described in the following work:
@article{carroll2019utility,
title={On the utility of learning about humans for human-ai coordination},
author={Carroll, Micah and Shah, Rohin and Ho, Mark K and Griffiths, Tom and Seshia, Sanjit and Abbeel, Pieter and Dragan, Anca},
journal={Advances in neural information processing systems},
volume={32},
year={2019}
}
To Do
[] Clean up unused code (Randomised starts)
[] Update dynamics to match latest version of Overcooked-AI