Evoluting Life

This application is an experiment to train robots to travel without colliding. To achieve this, it uses a Genetic Algorithm to train a Recurrent Neural Network. The fitness of a robot is its age whihc in turn is indirectly determined by traveling a distance as large as possible within a limited amout of time and with no collisions. Robots can be run in a simulation as well as on a raspberry-pi.

Build

To build this application using a Terminal, navigate to the root of this project. Make sure you have installed Maven and type:

mvn clean package

Running the simulation

To run this application as a simulation using a Terminal, navigate to the target sub folder and type:

java -jar evoluting-life-java-<version>-jar-with-dependencies.jar

Where the -placeholder needs to be replaced by the current version.

Running on a Raspberry-Pi

Prerequisites

pi4j

To control the pins on your Raspberry-Pi, this application uses the pi4j library. The library needs to be installed on your Raspberry-Pi. You can find instructions here: Install pi4j

ServoBlaster

To control the servo that holds and turns the sensor, the application uses ServoBlaster. This repository contains a servoBlaster.sh which should be started on the Raspberry-Pi before starting the application.

runRobot

To start the application in a convenient way, it is recommended to copy the runRobot.sh file in this repository to your Raspberry-Pi.

The application

Also copy the application 'evoluting-life-java--jar-with-dependencies.jar' that was previously build to your Raspberry-Pi.

Pin setup

The pin numbering of wiringpi wil be used in this description and in code. Visit Raspberry Pi Pinout for reference. The following pins are used for running the application on the Raspberry-Pi:

• WiringPi 07 : Servo (output)
• WiringPi 10 : Sensor echo (input)
• WiringPi 14 : Sensor trigger (output)
• WiringPi 23 : Left motor forward (output, hard PWM)
• WiringPi 25 : Left motor reverse (output, soft PWM)
• WiringPi 26 : Right motor forward (output, hard PWM)
• WiringPi 28 : Right motor reverse (output, soft PWM)
• WiringPi 29 : Led (output)

Structure of the code

The green colored boxes represent classes that are part of the core and are used in both simulation and on the Raspberry-Pi. The orange colored boxed represent classes that are used for the simulation only. The grey colored boxes represent classes that are used on the Raspberry-Pi only. The red lines represent a dependency that is used to report velocity, position, collision etc. to the SimRobot class.

Simulation physics

A robot has three motors: two motors to move around and one motor for turning the sensor. When a motor accelerates to move forward, backward or to turn the sensor, the motors use energy of the robot.

When a robot is moving or turning its sensor, it experiences drag. So without applying a force to accelerate, any movement will come to a halt.

Health of a robot

The health of a robot is determined by actions that cost energy and actions that produce energy. When a robot has zero or less energy, the robot is considered 'dead' and is replaced by the GA.

Time, acceleration and collisions cost energy. Travelling distance produces energy.

Each time tick costs some amount of energy, starting with a relative low number. When a robot from the population has reached its maximum age, the cost per time tick is increased by a small amount, so that robots will have to travel a larger distance per time tick in order to keep their health.

Recurrent Neural Networks

The robots are equiped with a Neural Network that contains Gated Recurrent Units. For more information on GRUs, see Understanding LSTM Networks by Christopher Olah.