Reinforcement learning (RL) has become a new paradigm for solving complex decision making problems, especially for robotic control and game AI. Human-in-the-loop learning focuses on human involvement during AI models training or deployment, so that AI technology can satisfy human preferences and benefit the whole human society.

This exhibit as part of Bringing the Human to the Artificial shows a recent attempt on training capable agents for human-AI coordination, lead by Yali Du, Ziyan Wang and Xingzhou Lou. They train adaptive agents for coordinating with humans without the need for human data, and the agents are deployed to play Overcooked with humans. You can play the game in the exhibition.

Instructions for playing Overcooked

You will be playing a cooking game. You will play one of two chefs in a restaurant that serves onion soups. The other chef is an AI agent.

Below is an example of what a level of the game looks like, with objects labelled.

Movement and interactions

Use the arrow keys and spacebar on the keyboard to play the game. 

You can move up, down, left, and right using the arrow keys, and interact with objects using the spacebar.

You can interact with objects by facing them and pressing spacebar. Here are some examples:

• Pick up ingredients (onions or tomatoes) by facing the ingredient area and pressing spacebar.
• To put an ingredient on an empty counter, you must be holding the ingredient then press spacebar.
• To put an ingredient in a pot, you need to hold an ingredient, face a pot that is not full, and press spacebar.
• To start cooking a coup, face a pot that is full and press spacebar. You cannot be holding anything.

You can't be in the same location in the kitchen as your partner.

Cooking

Once an ingredient is in the pot, you can begin cooking a soup by pressing spacebar. You can put up to 3 ingredients in a pot before cooking.

When a soup begins cooking, a red number will appear over the pot to denote the cook time. This number counts upward until the soup is done. The cook time is determined by its ingredients; each onion adds 20 to the cook time.

The base reward for any recipe is determined by its ingredients; each onion adds 21 points and each tomato adds 13 points.

Serving

Serve a soup by holding a plate, interacting with the cooked soup to get it in the bowl, then take it to a grey serving counter.