From Models of Cognition to Robot Control and Back Using Spiking Neural Networks.

With the recent advent of neuromorphic hardware there has been a corresponding rise in interest in spiking neural network models for the control of real-world artificial agents such as robots. Although models of cognitive mechanisms instantiated in spiking neural networks are nothing new, very few of them are translated onto real robot platforms. In this paper, we attempt such a translation: we implement an existing, biologically plausible model of reaching (the REACH model) demonstrated in 2D simulation on a UR5e robot arm. We are interested in particular in how well such a translation works since this has implications for similar exercises with a vast library of existing models of cognition. In this particular case, after extensions to operations in 3D and for the particular hardware used, we do find that the model is able to learn on the real platform as it did in the original simulation, albeit without reaching the same levels of performance.