Hardware evolution of analog circuits for in-situ robotic fault-recovery

We present a method for evolving and implementing artificial neural networks (ANNs) on field programmable analog arrays (FPAAs). These FPAAs offer the small size and low power usage desirable for space applications. We use two cascaded FPAAs to create a two layer ANN. Then, starting from a population of random settings for the network, we are able to evolve an effective controller for several different robot morphologies. We demonstrate the effectiveness of our method by evolving two types of ANN controllers: one for biped locomotion and one for restoration of mobility to a damaged quadruped. Both robots exhibit nonlinear properties, making them difficult to control. All candidate controllers are evaluated in hardware; no simulation is used.