Evolves a neural controller for walking locomotion. A physical model is defined via configuration file, and neural networks compete and evolve to see who can control it (with specified muscles and sensors) and best satisfy the objective function, which is currently just how far can it move in a fixed number of time steps.

Includes my own (almost complete) implementation of Ken Stanley's NEAT genetic algorithm for neural networks, which can be used separately from the walking code. The NEAT sub-directory includes test code for learning XOR and pole balancing (standard tasks mentioned in Stanley's thesis).

The physics simulation uses Box2D, so figure just a 2D rigid body, which makes two legged locomotion far easier (don't need to balance along z-axis). Limbs are specified in terms of boxes and circles, with simple revolute joints. Some form of walking motion is usually achieved, most likely due to the passive-dynamic nature of the muscle controls. Muscles are modeled as damped springs, and the neural controller can modify both the desired length of the muscle, along with the spring constant to control whether the muscle should be tense or loose.

The code depends on:

• boost (shared_ptr, function, mem_fn, bind)
• libconfig
• SDL, SDL_ttf, SDL_gfx
• Box2D v2.1.2

Optionally Intel Threading Building Blocks (TBB) can be used to parallelize the trivially parallel fitness calculation across the population each generation.

If these libraries are installed, and the locations in the Makefile are appropriate, then just running gnu make should be enough.

Two programs are created when make is run. legs simply tests the physical configuration and simulation, dropping whatever figure is given to it via -C configuration file, and watching it fall.

hopper runs the actual genetic algorithm, optionally displaying fittest member every 10 generations. Command line parameters are:

• -C configuration file
• -N number of generations (default 1000)
• -V if included, display simulation of fittest member every 10 generations

The rest of the many (many) algorithm parameters are specified in the configuration file, along with the specification of the walker.

The configuration file is formatted to be read by libconfig, which has a simple and hopefully fairly obvious structure and formatting. The config file is broken up into sections:

• global for NEAT genetic algorithm parameters
• limbs for specifying a list of physical limbs of the creature
• joints join limbs by name (currently just revolute type joints)
• muscles specify a list of joints that attach two limbs
• shapes name points on the limbs to be referenced in sensors
• sensors specify input that translate body position to input to the network. There are currently 3 types of sensors: Joint angle, Height of shape, Limb angle.

Three example configuration files are included:

• hopper.cfg: one legged hopper
• walker.cfg: two legged walker, now with feet
• kanga.cfg: fanciful one legged, 2-D kangaroo with big foot and tail

• Better statistics gathering during simulation
• Remove stagnating species
• More Sensors (force, ground location)
• Energy consumption based objective function
• Ramped objective for more complicated tasks