This project is based on Dynamic Field Theory (DFT). DFT is a mathematical formulation of neural behavior at the population level, used to model various perceptual, motor, and cognitive functions.
DFT is here used to form attractor patterns representing a sequence of actions towards a goal. These dynamic plans can handle temporary occlusions, very high level of noise and continuously update as conditions in the environment change.
Two example implementations are provided with this project:
- A disembodied agent moving around on a 2D surface, finding its way from start to goal. This simulation runs completely within MATLAB.
- A controller for an E-Puck miniature robot, simulated in Webots. This example depends on MATLAB, Webots, and Python.
This implementation has been used to evaluate the SPA framework, please refer to Billing et al. (submitted) for more details and background on the theory.
First of all, we recommend that you download SPA and other dependent projects by checking out the code directly from the repository. For this, you'll need Mercurial installed.
The main parts of SPA is implemented in MATLAB. SPA is tested on 64bit MATLAB 2013a, 2013b, and 2014a, under Linux and OS X, but should work also on other versions of MATLAB. Please refer to the Cosivina documentation for further details on compatibility with other MATLAB versions.
Some parts of SPA also requires the following software:
- Cyberbotics Webots Pro 7.4 (Unfortunately, the free version of Webots does not work)
- Python
Note that there are some limitations in the compatibility between MATLAB and Webots, please refer to the Webots Documentation for details.
Stary by checking out the source code for SPA into a folder of your choice, for example ~/Documents/MATLAB/Spa, using the following command:
hg clone https://bitbucket.org/interactionlab/spa ~/Documents/MATLAB/Spa
SPA depends on Cosivina. By default, SPA will look for Cosivina in ../Cosivina/, relative to the location where you checked out SPA. We therefore recommend that you check out Cosivina into ~/Documents/MATLAB/Cosivina:
hg clone https://bitbucket.org/sschneegans/cosivina ~/Documents/MATLAB/Cosivina
You may also want to install JSONlab. This is not required for the core functionality of SPA but used by Cosivina to store user preferences. Please refer to the Cosivina documentation for details.
Now start MATLAB, navigate to the SPA root directory and and initiate the framework by running setpath:
>> setpath
If you use a custom location for Cosivina, specify that as the first argument to setpath:
>> setpath /My/Cosivina/Dir
The setpath script registers temporary directories in the MATLAB path, and will need to be reexecuted after you restart MATLAB.
This example implements a disembodied agent moving around on a 2D surface, finding its way from start to goal. This simulation runs completely within MATLAB and may be executed with the following command:
>> launcherContinuousSearch
This launcher should bring up a window with six plots and some controls at the bottom. For further explanations of this simulation, please refer to the SpaDocumentation.pdf provided with this package.
This example runs through Webots and comprise a simulated E-Puck robot finding its way from a random start location to the goal, in a simple maze environment. The example is executed by opening the Webots world file e-puck-fieldsearch.wbt located in ~/Documents/MATLAB/Spa/webots/epuck/worlds, (assuming that you placed SPA in ~/Documents/MATLAB/Spa.
When the simulation runs you'll see a plot window similar to the one in Example 1. For further explanations of this simulation, please refer to the SpaDocumentation.pdf provided with this package.
References
E. A. Billing, Robert Lowe, and Yulia Sandmirskaya. Simultaneous Planning and Action: Neural-dynamic Sequencing. Submitted to Paladyn, Journal of Behavioral Robotics. Versita. 2014.