Memory Based Tactile Perception for Task Relevant Action Selection in Robots

Developments in robotic tactile sensing, including the transduction of physical forces analogous to biological mechanoreception, have significantly enhanced robots' abilities to grasp, manipulate objects, and explore their environments. The integration of touch has driven advancements in perception and memory models capable of processing and actively storing tactile information. Emulating biological memory systems holds great promise for advancing autonomous systems by enabling robots to store and utilise memories from experiences, thereby performing complex tasks using contextual sensorimotor information.

This work investigates the use of non-linear probabilistic dimensionality reduction techniques to abstract fundamental functional properties of memory, such as compression, pattern separation, and pattern completion. Additionally, a multiview learning approach is proposed as a unified model for memory and perception of tactile properties critical to the execution of sensorimotor tasks. Improvements in the predictive capabilities for geometric and spatial quantities are demonstrated through the application of hierarchical non-parametric probabilistic models.

The results of this work highlight the robustness of probabilistic models in representing memory and perception of tactile properties, even with relatively small datasets. Furthermore, their efficacy in enabling the execution of sensorimotor tasks featuring active touch sensing underscores their potential for advancing robotic tactile perception.