PUB - Publikationen an der Universität Bielefeld

In this study, we expand our work on Echo State Network (ESN) based forward models for predicting the expected sensor reading of an insect- inspired active tactile sensor system on a mobile robot. Inputs to the forward model are the motor commands which set the positions of both antennae, and a local proprioceptive signal which measures the vibra tions of the robot platform. The model was successfully used to detect tactile events on each antenna while the robot, is either moved on a straight or curved path with bumps on the floor, or moved with spiked wheels. The latter scenario was chosen to estimate the performance of an active tactile system mounted to a walking robot with rhythmic impacts caused by the legs touching down. By simulated lesion experiments, we found that the local proprioceptive signal plays a major role in discriminating disturbances from tactile events in the proposed forward model. We conclude that such an ESN-based forward model is suitable for prediction and elimination of self-induced stimulation of sensors, in particular of mechanosensors.