DinoAIChallenge

This is an AI challenge based on a clone of the game when the Chromium Browser detects that you are offline... Anther clone of that game is here. The challenge or opportunity is to develop/train an AI that can play this game.

The game in this repository exposes an interface to the in game data via a UDP connection with JSON messages

The Idea and other inspirations

For learning the basics of AI a lot of old games have been rewritten with interfaces to be ready to be experimented with. For example in the OpenAI Gym

However there is a very tight coupling between the game, especialy the simulation step, and the AI algorithms. I want to try to (forecefully) decouple it. The game will run in normal speed and transmit the information via a UDP connection and will receive the commands also via this connections. The goal is thus to have a forced "real-time" training and gameplay to be a little bit closer to the reality e.g. robotics, where a simulation might not be available and the real physical process to be learned can not be stopped or waited for.

A similar attempt especially with this game was made and is available on youtube - the code is also here on [github] (https://github.com/Code-Bullet/Google-Chrome-Dino-Game-AI) Also here the AI and game are also coupled in Proccessing.

The clone of the Dino Game

The pure game

The clone of the game is written in Processing and is in the Folder DinoAIProcessing. The game if fully playable.

Key	Action
space	jump
down arrow	duck
r	restart
i	toggle score display
d	toggle debug view

It can be loaded in processing or in the latestes release here

The UDP interface

The game listens for a udp message in a certain format on port 25001. Once a connection is made the game will respond on this connection and send its data with each frame. A python example is in this repository

Inital trigger message

To trigger the program to send out UDP message it needs to be triggered. This can be done by sending anything to port 25001 on the machine where the dino game is running

Message format for receving data from the game

Once the game is triggered by the initial message it will respond with this telemetry.

{   "level" : 1,
    "height_obstacle" : 0,
    "instances" : 3,
    "distance_obstacle" : 127,
    "gameover": False,
    "dinos" : [{"id": 0, "score": 13, "player_height": 110}, {"id": 1, "score": 13, "player_height": 57.20000076293945}, {"id": 2, "score": 13, "player_height": 97}]}

The fields in the JSON message

Field	Meaning
level	The current level
height_obstacle	The height of the next obstacle (the graphics y-coodinate 0 top - 150 high)
instances	The number of independent/simultaneous dinos in the game
distance_obstacle	The distance to the next obstacle
dinos	The array of all dino instances with specific information

The specific dino message

Field	Meaning
id	The id of the dino instance
score	The score of this dino (will remain constant if dino dies)
player_height	The heigth of the dino (important when jumping)
alive	True if the dino is still alive

Message format for sending to the game

A valid message sent to the game contains two mandatory fields: command and num_instances command The command can be left empty and the num_instances determine how many dinos will be in the game simultaneously (max. 20).

    msg = {
        "command" : " ",
        "num_instances" : 3
    }

If the command is left empty the game will continue normally. The num_instances describes how many controllable dinos will be in the game simultaneously. Those dinos are totally independed from each other. The idea behind is to ease the training of heuristic or genetic algorithms and to test the differet mutations at the same time. The message as displayed is essential and if the num_instances is changed the game will restart.

Command	Description
restart	will issue the game to restart

Commanding the dino

To command the instance of the dino the message has to contain an array called dinos with an objecy containing the fields action and dino_instance. The dino_instances addresses the dino and the action can be one of the following:

Action	Description
smalljump	the dino will do a small jump
bigjump	the dino will do a big jump
duck	the dino will duck

Example for a big jump

{
    "command" : " ",
    "num_instances" : 3
    "dinos" : [{"action" : "bigjump", "dino_instance" : 0}]
}

Interface Example

In the example is a python program to use the socket to communicate with the game and

TODO

Scale of the obstacles
Tuning of the obstacle distance so that the smalljump is required
messages the distance between the obstacles

kruegerrobotics / DinoAIChallenge