Hardware architectures for neural processing systems for bio-inspired robot control

This PhD Thesis is focussed in the study of spiking neural networks. In this framework the presented work presents different hardware architectures that are implemented in reconfigurable devices (FPGAs). Different approaches are proposed adopting time¬driven or alternatively event-driven processing schemes. The work presents alternative control approaches in the field of robotics and studies computing architectures for the simulation of massive spiking neural networks of millions of neurons processing sensorimotor information in real-time. These proposed approaches have been implemented in two hybrid Hardware/Software platforms with different levels of autonomy of the hardware (stand-alone and co-processing strategy) with respect to the software modules (in a PC as a host computer) that simulated in real-time these large scale networks. In a second stage, this Thesis focuses on experiments with real-robots, as validation methodology of the control neural networks under study. The choice of working with real robots instead of simulated ones in motivated by the difficulty of describing in a realistic way the interaction with the real-world in a simulated framework. Therefore, the work here also adopts the "Embodiment concept" which stresses the necessity of having a physical body as learning mechanism for the knowledge emergence generation. In this field, the Thesis describes two robotic platforms built and adapted for being controlled by spiking neural systems. The obtained results show that imitating in more or less detail the biology is feasible building neural circuits which represent valid alternatives to be considered for control of biomorphic robots with complex physical structures.