What you choose to see is what you get: An experiment with learnt sensory modulation in a robotic foraging task

Abstract

In evolutionary robotics, the mapping from raw sensory input to neural network input is typically decided by the experimenter or encoded in the genome. Either way, the mapping remains fixed throughout a robot’s lifetime. Inspired by biological sensory organs and the mammalian brain’s capacity for selective attention, we evaluate an alternative approach in which a robot has active, real-time control over the mapping from sensory input to neural network input. We augment the neural controllers with additional output neurons that control key sensory parameters and evolve solutions for a single-robot foraging task. The results show that the capacity to control themapping from raw input to neural network input is exploited by evolution and leads to novel solutions with higher fitness compared to traditional approaches.