The code for my master's thesis "Evolving steerable pneumatic soft robots".

This code uses the CPPN-NEAT algorithm to simultaneously evolve the body and controller of voxel-based soft robots with realistic pneumatic actuation and sensing. The two tasks supported are simple gait learning and directed locomotion. Videos of some of the resulting robots can be seen here. For more information, see my thesis.

Dependencies

Using voxelyze for simulation and raylib (compiled statically) for visualization. We use the evolutionary algorithms module neat-python, and tiny-dnn for the NN controllers.

Installation

The folder structure should be like this:

cpu_voxelyze
|- pneuvox (this repo)
|- raylib
|- tiny-dnn
|- Voxelyze

Compiling tiny-dnn

Tiny-dnn is header-only so doesn't need to be compiled separately.

Compiling Voxelyze

Use my personal fork (https://github.com/th555/voxelyze) on the branch fixext, then just run make

Compiling Raylib

Skip this step on the DAS-6 cluster!

Run make in its directory, then copy raylib.h and raymath.h from raylib/src to raylib/include and move libraylib.a from raylib/src to raylib/lib

Compiling Pneuvox locally

Normal make builds a local version, requiring Raylib for visualization.

Compiling Pneuvox on DAS-6

make das6 builds a version optimized for the DAS-6 cluster, with compiler options for best performance on its CPUs, and no visualization. You can also add -march=znver2 to the makefile.

Installing Python dependencies

Make sure you have numpy and scipy, on DAS-6 this is

module load py3-numpy
module load py3-scipy

Install other deps:

pip install -r evo/requirements.txt

or on the das:

python3 -m pip install --user -r evo/requirements.txt

Viewing some robots

For a quick check if the installation/compilation succeeded, try loading some of my previously evolved robots:

python evo/replay.py bestbots/500_mdv2_open

Viewer controls

RMB: pan alt+RMB: rotate view ctrl+alt+RMB: zoom with mouse scroll wheel: zoom R: reset simulation esc: exit

Running on DAS-6

Start every session on the das with module load prun.

Basic prun: prun -v -np 1 -1 python3 -u evo/run.py experiment_name

Restoring an evolution run at a generation checkpoint: prun -v -np 1 -1 python3 -u evo/run.py restore saved_data/10x10/checkpoint_23

Long running job, you can safely log off ssh (not allowed during working hours!): nohup 2>&1 prun -v -np 1 -1 -asocial -t 09:00:00 python3 -u evo/run.py experiment_name > experiment_name.out &

(keep an eye on it with watch -n 5 tail -n 40 experiment_name.out)

(though of course it's better to use a screen instead of nohup)

Retrieving your results

To transfer files from the DAS, via the stepping-stone server, to your local computer: scp [email protected] [email protected]:cpu_voxelyze/pneuvox/exp_results/results.tar.gz exp_results/results.tar.gz

If it complains ("received message too long..."), try putting this on top of bashrc on the das6:

# If not running interactively, don't do anything and return early
[[ $- == *i* ]] || return