Autonomous Locomotion Control for a Snake-like Robot with a Dynamic Vision Sensor and a Spiking Neural Network

The repository for the Bachelor's Thesis in Informatics 'Autonomous Locomotion Control for a Snake-like Robot with a Dynamic Vision Sensor and a Spiking Neural Network' by Christoph Clement under the supervision of Zhenshan Bing, M.Sc. at the TU Munich.

• controller: Contains the python module for the SNN controlled Locomotion of the snake and plotting scripts for the training Figures.
• data: Contains h5f, csv and json files of the training and testing sessions
• plots: Contains plotting scripts for the testing Figures and other Figures from the thesis and the actual testing and training Figures.
• V-REP_scenarios: Contains the different V-REP scenarios.

Abstract

How does the brain process vast amounts of information with such great energy efficiency? Compared to modern Artificial Neural Networks (ANNs), the energy consumption of the brain is ten times lower - (20W) in comparison to (200W). On top of that, we are still far away from imitating the brain's capabilities with ANNs.

A solution to this question might come from the third generation of ANNs - Spiking Neural Networks (SNNs) that mimic the underlying mechanism of the brain better than current ANNs. They incorporate spatial and temporal information into their calculations leading them to be more computationally and energy efficient. Especially in combination with Dynamic Vision Sensors (DVSs), they are an excellent fit for autonomous robots where energy efficiency and fast real-time computations are critical success factors.

In this work, a SNN controller using the Reward-modulated Spike-Timing-Dependent-Plasticity learning rule for the autonomous locomotion control of a snake-like robot is implemented. The controller is trained in a maze environment and demonstrates the ability to cope with new situations in the form of different wall heights and maze angles during testing. In a real-world application, such a robot could be deployed in areas with uneven terrain like in a collapsed factory building.

Get Started

Prerequisites

• Ubuntu 16.04 LTS
• Python 2.7
• ROS Kinetic 1.12.13
• V-REP 3.4.0
• NEST 2.14.0

Installation

1. Install Ubuntu 16.04 LTS
2. Install ROS Kinetic

• Follow this guide: https://wiki.ros.org/kinetic/Installation/Ubuntu
• Make sure to use the Desktop-Full Install in step 1.4

  sudo apt-get install ros-kinetic-desktop-full
  

3. Install V-REP 3.4.0

• Download

  cd ~/Downloads
  wget http://coppeliarobotics.com/files/V-REP_PRO_EDU_V3_4_0_Linux.tar.gz
  cd ~
  tar zvxf ~/Downloads/V-REP_PRO_EDU_V3_4_0_Linux.tar.gz
  

• Set the VREP_ROOT variable in ~/.bashrc

  echo 'export VREP_ROOT="$HOME/V-REP_PRO_EDU_V3_4_0_Linux"' >> ~/.bashrc
  source ~/.bashrc
  

4. Install some required Ubuntu packages

sudo apt install git cmake python-tempita python-catkin-tools python-lxml default-jre

5. Install saxon

• Download

    cd ~/Downloads
    wget http://downloads.sourceforge.net/project/saxon/Saxon-HE/9.7/SaxonHE9-7-0-8J.zip
    cd ~
    mkdir -p saxon/bin
    cd saxon
    unzip ~/Downloads/SaxonHE9-7-0-8J.zip
    echo -e '#!/bin/sh\njava -jar "`dirname "$0"`/../saxon9he.jar" "$@"' > bin/saxon
    chmod a+x bin/saxon
  

• Update PATH env var with the location of saxon executable

    echo 'export PATH="$PATH:$HOME/saxon/bin"' >> ~/.bashrc
    source ~/.bashrc
  

6. Install v_repStubsGen (https://github.com/CoppeliaRobotics/v_repStubsGen)

  mkdir ~/python-packages
  cd ~/python-packages
  git clone --recursive https://github.com/CoppeliaRobotics/v_repStubsGen
  echo 'export PYTHONPATH="$PYTHONPATH:$HOME/python-packages"' >> ~/.bashrc
  source ~/.bashrc

7. Setup catkin workspace

  echo 'export PYTHONPATH="$PYTHONPATH:/usr/lib/python2.7/dist-packages"' >> ~/.bashrc
  source ~/.bashrc
  rm -rf ~/catkin_ws
  mkdir -p ~/catkin_ws/src
  cd ~/catkin_ws/src
  catkin_init_workspace
  cd ..
  catkin build
  source devel/setup.bash

• Copy ~/V-REP_PRO_EDU_V3_4_0_Linux/programming/ros_packages/v_repExtRosInterface to ~/catkin_ws/src

    catkin build
  

• Copy /home/christoph/catkin_ws/devel/lib/libv_repExtRosInterface.so to ~/V-REP_PRO_EDU_V3_4_0_Linux

8. Install NEST 2.14.0 from http://www.nest-simulator.org/

Start a simulation

• Set parameters (training/ testing length, reward factor, resolution, etc.) in ./controller/parameters.py
• Start a training session by running the following command from the ./controller/ folder. After the training, close the terminal that opened up.

   training.py
  

  (Uses scenario_eight_2_0 as the default training scenario. If you want to train on another scenario, run the following command:)

  training.py scenario_eight_0_5
  

• Start a testing session by running the following command from the ./controller/ folder. After the training, close the terminal that opened up.

   controller.py scenario_eight_0_5
  

  (Uses scenario_eight_2_0 as the default training scenario. If you want to train on another scenario, run the following command:)