Escape AI

AI learns escape a room.

This is a rust based implementation of a genetic algorithm and reinforcement learning simulation that trains an AI named Zoe to escape from five rooms of increasing difficulty. The simulation is built using the macroquad library.

Demo

Usage

• Clone the repo

  git clone git@github.com:bones-ai/rust-escape-ai.git
  cd rust-escape-ai
  

• Run the simulation

  cargo run --release
  

• To update the simulation configurations use the configs file located at src/configs.rs

Create custom rooms

• Tiled map editor is used for building the 2d levels
• To create custom levels and play around with the AI, use tile layers with names player, keys, door, walls, spikes, enemies and background. I suggest taking a look at how other maps are built to understand how things are to be setup.

Configurations

• The project config file is located at src/configs.rs
• The game rooms have to be manually configured in the config file.

Inputs

• Spacebar - Pause/Unpause the simulation
• Tab - Show/hide the egui control menu
• r - Restart the simulation
• Backspace - Slow mode
• Backslash - Enable AI, start the AI training process (Use this to play the game yourself, using keyboard inputs wasd)
• Right Shift - Run at 5x speed
• Mouse wheel - Zoom in/out
• Mouse wheel with ctrl - Large zoom in/out
• Right mouse click drag - Pan through the world
• To draw all training AI agents, use the egui menu, then select Show Multi. Caution, if you have too many game rooms being simulation, this could freeze your machine.

Assets

• https://fisherg.itch.io/micro-asset-pack

If Zoe isn't able to solve a room

• Try restarting (using shortcut r). It's impossible to get the same results as in the youtube video as every simulation run is random. Sometimes Zoe might get stuck (or spend a lot of time/steps) in a region, letting the simulation run for a longer is an option (but it'll take a long time).
• Update NUM_FRAMES to a larger value. Once the room is solved, Zoe will try to solve the room more efficiently (i.e taking less steps)
• Run the simulation at 5x (by using shortcut Right Shift). Some rooms can take a lot of generations to solve.
• Disable the drawings in the egui menu to speed up the training process