ANNarchy - iCub: An Interface for Easy Interaction between Neural Network Models and the iCub Robot

Abstract

Models and algorithms developed in the computational neuroscience or machine learning domain are able to solve more and more complex tasks. At the same time there is an increasing demand to test these algorithms in virtual or real robot platforms. The Python programming language is commonly used as interface language for machine learning tools (e. g. Keras) or neural simulators (e. g. ANNarchy, Brian2, PyNEST) as the language is easy to use and therefore allowing a quick development of algorithms or models. Robotic platforms, as the here investigated iCub, are using the YARP middleware written in C/C++ as it offers low-level access and high performance. Hence, to implement a closed loop interaction between a model and the robot one needs to connect these two languages without keeping the modeler in the loop. Therefore, we present an ANNarchy-iCub interface providing the user a Python API to a selected set of operations offered by the iCub. In contrast to the YARP Python API, the proposed interface is specialized for the requirements of neural network models and encapsulates the interaction with the robot. The interface capabilities are shown by a simple demonstration where the iCub moves his arm along a given trajectory which can be interrupted by an obstacle. We also provide some insight into the performance considering two important tasks: image processing and joint motion, by comparing our interface implementation with the Python API provided for YARP.