Tactile and proprioceptive online learning in robotic contour following

Purposive and systematic movements are required for the exploration of tactile properties. Obtaining precise spatial details of the shape of an object with tactile data requires a dynamic edge following exploratory procedure. The contour following task relies on the perception of the angle and position of the sensor relative to the edge of the object. The perceived angle determines the direction of exploratory actions, and the position indicates the location relative to the edge for placing the sensor where the angle tends to be perceived more accurately. Differences in the consistency of the acquired tactile data during the execution of the task might induce inaccuracies in the predictions of the sensor model, and therefore impact on the enactment of active and exploratory movements. This work examines the influence of integrating information from robot proprioception to assess the accuracy of a Bayesian model and update its parameters to enhance the perception of angle and position of the sensor. The incorporation of proprioceptive information achieves an increased number of task completions relative to performing the task with a model trained with tactile data collected offline. Studies in biological touch suggest that tactile and proprioceptive information contribute synergistically to the perception of geometric properties and control of the sensory apparatus; this work proposes a method for the improvement of perception of the magnitudes required to actively follow the contour of an object under the presence of variability in the acquired tactile data