A forward model for an active tactile sensor using Echo State Networks

Here, we introduce a forward model designed for predicting the expected reading of a bionic tactile sensor (antenna) mounted onto a wheeled robot. The model was used to distinguish self-generated stimulation from true tactile events to the antenna. An Echo State Network (ESN), a special type of recurrent neural network which is suitable for chaotic time series prediction, is used to implement the forward model. Inputs to the ESN are the motor command which sets the position of the antenna, and a local proprioceptive signal which measures the acceleration of the robot platform. The model can successfully be used to detect a tactile contact on the antenna while the robot is moving along a path with obstacles. Such forward models are good candidates to be used in neural yet simple way to eliminate self-stimulation of sensors of other modalities due to ego-motion.

I. Introduction

In animals, near-range exploration, especially the active tactile sense is often of great importance; many insects actively use their antennae as tactile sensors for obstacle localization, orientation, pattern recognition and even for communication [1]. In the same manner, mammals like cats or rats use active whisker movements to detect and scan objects in the vicinity of the body. Compared to vision based sensors, the tactile sense is independent of light conditions, it works at day and night. Here we use an active bionic tactile sensor inspired by the antenna of the stick insect, Carausius morosus and it was codeveloped by Fraunhofer IFF, Magdeburg and the University of Bielefeld [2]. The sensor is capable of near-range tactile localization and material classification of contacted objects [3]. It uses an acceleration sensor mounted to the tip of an otherwise unsensorised probe. Contact distance is determined by using the peak frequency of the damped oscillations of the probe and the damping properties are used for material classification. However, if the tactile system is attached to a mobile platform such as a walking robot, the sensor signal will be corrupted by the self-induced movements. For instance, the antenna is in continuoues motion (oscillating) and that will directly affect the mechanosensory signal. Additionally, when the robot is moving on a rough terrain, the vibrations of the platform can propagate to the tactile system. Such a corrupted signal would impair the tactile performance. Hence, in the present study, we focus on designing an internal forward model to estimate the expected sensor output of the active tactile sensor, especially when the system is in motion. We then use the model in a prototypical example to detect tactile contact events on the antenna. Eventually, the sensor is to be used on a walking robot where it will be exposed to rhythmic impacts caused by the feet making ground contact. Here, we simulate such impacts by having the wheeled robot drive across bumps on the floor.