Morphology and Learning - A Case Study on Whiskers

Abstract

In many animal species, whiskers are an important sensory modality. Rats for example navigate, explore and recognize objects by actively whisking them, seals even use them to hunt in murky water. Whiskers also make an excellent model system for biologists due to their well defined anatomy. Recently, whiskers have also attracted an increasing interest by roboticists. We have built an artificial whisker system and mounted it on a robot. In this study, we investigate the role of whisker morphology on the behavior of a robot. For a robot to be able to navigate it has to react to obstacles appropriately yet in a flexible manner. For basic behaviors, associations between reflexes and stimuli can be learned using a model of classical conditioning, the distributed adaptive control (DAC) architecture. This model of associative learning is used to correlate whisker signals with collision signals from the robot's infrared sensors for whisker-based obstacle avoidance. The morphology of the whiskers is varied to test its influence on learning and behavior. We find that the performance of the robot is clearly influenced by its morphology. Furthermore, the weights learned under different morphologies vary according to the morphology. Finally, from the performance of the robot we can speculate about a suitable weight matrix for this task.