lloydtao / ecmm409-quadruped

Training and evolution of a virtual four-legged robot.

ecmm409-quadruped

Training and evolution of a virtual two-legged robot.

For the Virtual Creatures Competition 2022 (https://virtualcreatures.github.io/).

Demo

https://www.youtube.com/watch?v=7wZyJ0yNZHA

Purpose

This project implements a 3D computer simulation of a two-legged walker, built using pybullet and pyrosim.

The demonstration is followed by a showcase of the genetic algorithm optimising the walker over 6 generations. For each generation, the fitness score of each member is measured in a dynamic simulation. The fittest members of the population are selected, and children are produced based on each pair. Children have a small chance to mutate, in order to escape local maxima.

The end goal is a child that can travel as far as possible in the fixed time that the simulation runs for, which is 2160 frames.

Installation

Make sure you're running Python 3.9 with pip3 installed.

Set up and activate a virtual environment:

$ python -m venv '.env'

To activate:

$ source .env/bin/activate          # Linux
$ source .env/Scripts/activate      # Windows

Install the required pip packages to your environment:

$ pip install -r requirements.txt

Install pybullet (bullet3) manually, as the pip distribution is buggy:

$ git clone https://github.com/$ bulletphysics/bullet3
$ cd bullet3
$ python setup.py build
$ python setup.py install

Usage

All files should be run from the root directory (i.e. ecmm409-quadruped/).

To run the full evolutionary simulation:

$ python quadruped/evolution.py

To generate an example creature for the manual simulation:

$ python quadruped/generation.py

To run the manual simulation with the generated creature:

$ python quadruped/simulation.py