Implicit policies for deformable object manipulation with arbitrary start and end states: A novel evolutionary approach

We present a novel method for training (evolving) fully convolutional neural networks (CNNs) for deformable object manipulation. Instead of using a weight update rule, we evolve an ensemble of compositional pattern generating networks (CPPNs) by means of a genetic algorithm (GA). These ensembles generate the convolutional kernels that comprise the CNN. This allows the GA to search for fit kernels in the space of spatial 2D patterns, which allows for evolution of larger networks than is feasible with direct evolution of the connection weight vector. We apply this method to a thread manipulation task, in an attempt to let the CNN grasp the deformation dynamics of the thread. We report results both on single-manipulation tasks, and on tasks requiring a sequence of two manipulations, establishing the viability of this approach.