Skip to content

A reinforcement learning environment for self-driving cars in the browser.

Notifications You must be signed in to change notification settings

thibo73800/metacar

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

75 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Metacar: A reinforcement learning environment for self-driving cars in the browser.

Contributions welcome

Metacar is a 2D reinforcement learning environment for autonomous vehicles running in the browser. The project aims to let reinforcement learning be more accessible to everyone through solving fun problems. Metacar comes with a set of a predefined levels, some harder to address than others. More levels and possibile scenarios will be added soon (pedestrian, bikes...). Furthermore, the library let you create your own levels and personalize the environment to create your desired scenario.

If you want to be part of the project, whether to implement features in the environment or demonstrate algorithms, feel free to join the slack channel to ask questions and talk about all your fantastic ideas!

To start developing with metacar check out the Documentation and the API Reference

You can also take a look at the online demo.

Documentation

Table of contents:

  1. Getting Started
  2. Your first environment
  3. Interact with the environment
  4. Custom the environment
  5. Edit your own level

Getting started

Installing Metacar

You can use Metacar with a direct link in your HTML file or install it from NPM. However, metacar is based on Pixi.js: 4.7.1, then you need to include pixi.js as a global dependency in your HTML.

Script Tag

<!DOCTYPE html>
<html>
<head>
    <meta charset="utf-8" />
    <title>Metacar: Documentation</title>
    <script src="https://cdnjs.cloudflare.com/ajax/libs/pixi.js/4.7.1/pixi.min.js"></script>
</head>
<body>
    <script src="https://cdn.jsdelivr.net/combine/npm/[email protected],npm/[email protected]"></script>
</body>
</html>

OR Script Tag and NPM

npm i metacar
<!DOCTYPE html>
<html>
<head>
    <meta charset="utf-8" />
    <title>Metacar: Documentation</title>
    <script src="https://cdnjs.cloudflare.com/ajax/libs/pixi.js/4.7.1/pixi.min.js"></script>
</head>
<body>
    <script src="yourscript.js"></script>
</body>
</html>
import Metacar from "metacar";

Your first environment

Even if you can create your own level, Metacar comes up with a set of predefined levels accessible under metacar.level. Once your level is selected, you can use it to create a first environment.

// Select a level
const level = metacar.level.level1;
// Create the environement
const env = new metacar.env("env", level);
// Load it
env.load();

You also have to create the container in your HTML file.

<div id="env"></div>

(NOTE: metacar.env can be instantiated with an object or a string for the level parameters: doc API)

Wonderful! You just created your first metacar environment. You can take some time to play around with the arrow keys to move the car. The current collision system supports the detection of the following events:

  • Collisions with other vehicles.
  • Detection of the vehicles, ground, and road with the lidar.
  • Detection of the car going out track.

If you want to add new features to the detection system, you can consider contributing to the project :)

Interact with the environment

Action space

By default, the environment comes with a simple motion engine (How to change the motion engine ?) which lets you control the car with the arrow. Then, the actions are either UP, LEFT, RIGHT, DOWN, WAIT. Once the environment is loaded, you can take a look at the action space.

env.load().then(() => {
    console.log(env.actionSpace());
});
{
    type: "Discrete", // The number is discrete
    size: 1, // Only one number is expected
    range: [0, 1, 2, 3, 4] // The action can be either 0, 1, 2, 3 or 4
}

Play & Stop

Let's suppose your agent is already trained to move forward whatever happens (Fortunately we are in a simulation). Then you might want to test it in real time to see the result.

The quickest way to do so is to just ask the environment to call a given function at each loop turn.

env.load().then(() => {

    env.addEvent("play", () => {
        // Move forward
        const reward = env.step(0);
        // Log the reward
        console.log(reward);
    });

});

You should see a play button on the screen. On click, the car will move forward, and the reward should be positive as long as the car is on track, then negative when the car leaves the road.

To stop calling your function, you can add a stop button on the screen.

env.load().then(() => {

    env.addEvent("play", () => {
        // Move forward
        const reward = env.step(0);
        // Log the reward
        console.log(reward);
    });

    env.addEvent("stop", () => {
        console.log("The stop button have been pressed.");
    });

});

Train your agent

During the training, the environment is not rendering on the screen anymore. Once your training is finish you have to notify the environment by calling env.render(true) to render the environment again.

The state of the environment is made of four fields:

{
    a: number|undefined // Acceleration of the car (if any)
    lidar: number[][] // Lidar points values
    linear: number[] // The flatten lidar values + the current speed of the car
    steering: number|undefined // Steering angle of the car (if any)
}

Here is an example of simple training loop.

env.load().then(() => {

    env.addEvent("train", () => {
        for (let s=0; s < 100; s++){
            // Get the current state of the lidar
            const state = env.getState();
            // Move forward
            const reward = env.step(0);
        }
        // Log the reward
        env.render(true);
    });

});

Reset and shuffle env

To reset the environment, you can either call

    env.reset();

Or add a button to do it from the web page.

env.load().then(() => {

    env.addEvent("custom", () => {
        env.reset();
    });

});

You can also shuffle the position of vehicles (agent and other cars) on the map.

env.load().then(() => {

    env.addEvent("Shuffle only the agent", () => {
        env.shuffle({cars: false});
    });

    env.addEvent("Shuffle all", () => {
        env.shuffle();
    });

});

Custom the environement

!WARNING: The method presented in this section must be called BEFORE loading the environment.

Change the motion engine

There are two motion engine available: BasicMotion and ControlMotion.

BasicMotion

This is the default motion engine. Movement of the car is either up, down, left, right or wait. The car turns from a given angle for the left and right action.

You can change the parameters of the motion engine using the setMotion method.

env.setAgentMotion(metacar.motion.BasicMotion, {rotationStep: 0.25});
// Load the environment after having changed the properties.
env.load();

ControlMotion

The motion control is based on two continuous values for the throttle and steering angle of the car. Then the action is an array of two floating values. (see actionSpace)

env.setAgentMotion(metacar.motion.ControlMotion);
// Load the environment after having changed the properties.
env.load();

Change the lidar properties

There are four properties you can change. The number of points (pts) per line, the width and height of the area covered by the lidar and the position (pos) with respect to the car.

env.setAgentLidar({pts: 3, width: 1.5, height: 1.5, pos: 1});
// Load the environment after having changed the properties.
env.load();

Stop the others vehicles

You can choose to move or stop the other vehicles with env.carsMoving()

env.carsMoving(false);
// Load the environement after changing the propeties.
env.load();

Other methods

Load a file from your computer

This features can be useful to load the content of one file from your computer (the result of a trained model for instance).

env.load().then(() => {

    env.addEvent("load", (content) => {
        // Here the content of the loaded file.
        console.log(content);
    },{local: true});

});

Save a file on your computer

Also, you might want to save the result of a trained model on your computer.

env.save("content of my model", "model.metacar")

Add a custom event

env.addEvent() comes with a set of predefined event ("train", "play", "stop", "reset_env", "load") but you can also create custom event with an associated button on the page. Bellow, an example of custom event saving a file onClick.

env.load().then(() => {

    env.addEvent("My custom event", () => {
        env.save("content of my model", "model.metacar");
    });

});

Edit a new level

Only three lines are required to create the editor:

const level = metacar.level.level1;
var editor = new metacar.editor("env", level);
editor.load();

Left click on one item to select it. Then Left click on the map to set the item. Right click is used to remove an item.

Save the level

editor.load().then(() => {

    editor.addEvent("save", (content) => {
       // Save the content into the localstorage here or just
       // retrieve the downloaded json.
    }, {download: true, name: "mylevel.json"});

});