Reinforcement Learning 2D Humanoid Walkers
-"Play that funky music, robot"
- -
-
-
-
+
+
+
+
+
+
Reinforcement Learning 2D Humanoid Walkers
+"Play that funky music, robot"
+ + + - - -Learning
- This is an experiment to teach 2d humanoids to walk and to display their progress in the browser.
-
- This uses use reinforcement learning, a branch of machine learning target at controlling systems, to teach them to walk. Training is done offline in tensorflow.js. -
- Credits: The walker code is adapted from rednuht.org/genetic_walkers/, the DDPG code was adapted from github.com/thibo73800/metacar. + Teaching 2d humanoids to walk in your browser using reinforcement learning. You can view the progress at differen't stages of training with the buttons below. +
+ Technical: This uses use reinforcement learning to teach them + to walk. This is a branch of machine learning targeted at controlling systems such as system of limbs. Training is done offline in tensorflow.js. The environment is in box2d for javascript. The aglorithm is DDPG with prioritized experince replay. + + The dark outlines are when the agent grips the floor, since I found walking was difficult otherwise. The balls are to provide obstacles. +
+ Credits: The walker code is adapted from rednuht.org/genetic_walkers/, + the DDPG code was adapted from github.com/thibo73800/metacar. +
+ - This uses use reinforcement learning, a branch of machine learning target at controlling systems, to teach them to walk. Training is done offline in tensorflow.js. -
- Credits: The walker code is adapted from rednuht.org/genetic_walkers/, the DDPG code was adapted from github.com/thibo73800/metacar. + Teaching 2d humanoids to walk in your browser using reinforcement learning. You can view the progress at differen't stages of training with the buttons below. +
+ Technical: This uses use reinforcement learning to teach them + to walk. This is a branch of machine learning targeted at controlling systems such as system of limbs. Training is done offline in tensorflow.js. The environment is in box2d for javascript. The aglorithm is DDPG with prioritized experince replay. + + The dark outlines are when the agent grips the floor, since I found walking was difficult otherwise. The balls are to provide obstacles. +
+ Credits: The walker code is adapted from rednuht.org/genetic_walkers/, + the DDPG code was adapted from github.com/thibo73800/metacar. +
+
+
-
+
-
-
-
-
-
-
diff --git a/src/index.js b/src/index.js
index 9333102..053e1a5 100644
--- a/src/index.js
+++ b/src/index.js
@@ -2,6 +2,8 @@ const { Game } = require('./js/game')
window.b2 = require('./vendor/jsbox2d')
const config = require('./js/config')
+require('./css/walkers.css')
+
var game
window.game = game
diff --git a/webpack.config.js b/webpack.config.js
index 2da21cb..a4c4f30 100644
--- a/webpack.config.js
+++ b/webpack.config.js
@@ -1,5 +1,9 @@
const path = require('path');
const webpack = require('webpack')
+const HtmlWebpackPlugin = require('html-webpack-plugin')
+var CopyWebpackPlugin = require('copy-webpack-plugin')
+
+var ENV = process.env.NODE_ENV || "development";
module.exports = {
entry: {
@@ -7,13 +11,19 @@ module.exports = {
// lib: './src/vendor/index.js'
},
target: 'web',
- devtool: 'dev-source-map',
+ devtool: ENV==="development"? 'dev-source-map': null,
output: {
- path: path.join(__dirname),
- filename: 'dist/[name].bundle.js',
+ path: path.join(__dirname, 'dist'),
+ filename: '[name].bundle.js',
// path:'/dist',
},
- devServer: {
+ module: {
+ rules: [
+ {
+ test: /\.css$/,
+ use: [ 'style-loader', 'css-loader' ]
+ }
+ ]
},
optimization: {
splitChunks: {
@@ -39,6 +49,13 @@ module.exports = {
new webpack.DefinePlugin({
// A flag to disable node imports, and enable window/dom usage
WEB: true
- })
+ }),
+ new HtmlWebpackPlugin({ template: '!!html-loader!src/index.html' }),
+ new CopyWebpackPlugin([
+ { from: 'src/images', to: 'images' },
+ {from:'checkpoints',to:'checkpoints'}
+
+ ]),
+
]
};