AI Agent: Surviving a 3D Obstacle Course

An agent that uses policy iteration to run through a 3D Temple Run / Subway Surfers style environment

To see some of the thought process behind it, look in docs/

Setup

This assumes you have already cloned this repo. If not, run git clone https://github.com/ravishende/ai_agent_3d.git; cd ai_agent_3d first.

1. Create and run virtual environment python3 -m venv .venv source .venv/bin/activate
2. Install required packages python3 -m pip install -r requirements.txt

Run

Running the Agent

To see the agent play, run python3 run_rl_agent.py with some optional command line arguments:

• Ex: python3 run_rl_agent.py --length=15 --width=9 --difficulty=expert --visualize --quiet

• To change the output:

  • --visualize: Shows the agent running through the course in 3D
  • --quiet: Doesn't print verbose messages about the agent's process

• By default, maps will be generated randomly. Customize the map generation with the following:

  • --length=... (int): length of the randomly generated map
  • --width=... (int): width of the randomly generated map
  • --difficulty=... (str): difficulty use with options easy, medium, hard, or expert.
  • --trap_spawn_prob=... (float): frequency of traps in the map (from 0 to 1) - default is 0.3
  • --trap_death_prob=... (float): probability of a trap killing the player upon being activated (from 0 to 1) - default is 0.3

• If you prefer a pre-determined map, you can specify which map to use with --fixed --map=....

  • Ex: python3 play.py --fixed --map=3
  • Make sure to include --fixed if you don't want a randomly generated map.

• To change the 3D visualization of your game, specify the following parameters

  • --cubesize=... (float): sidelength of the visualized cubes (default is 2.0)
  • --spacing=... (float): space between each slice of cubes (default is 40.0)

A Note on Traps:

Traps add a nondeterministic element to the game. If a player ends up in the same column as a trap, the trap has a certain probability of ending the game. To disable traps, set --trap_spawn_prob=0.

While any randomly generated map is guaranteed to have a solution that avoids obstacles, it does not guarantee that the player can avoid all traps as well. if trap_spawn_prob=1, every grid that has more than one solution column will contain a trap. Grids with only one solution column will never contain a trap.

Traps add complexity to the game, forcing the agent to make informed decisions on whether to risk an early end for the chance of having a longer safe path in the future. If traps were entirely avoidable, they could be treated as obstacles, rendering them useless in terms of agent decision making.

Playing the Game

To be able to play through a text-based game and see it visualized, run python3 play.py with some optional command line arguments (any of the previously mentioned arguments in running the agent except --quiet).

• Ex: python3 play.py --width=5 --length=15 --difficulty=hard

To just play the text game, run python3 play_text.py

To just see the environment, run python3 env_3d.py