Smart Illuminated Touch: Integrated AI for Improved Optical Tomography-Based Soft Sensors

The development of tactile sensing is one of the most challenging, and indeed essential, tasks in robotic control. Tactile sensor skins based on optical tomography offer an attractive solution because of the absence of embedded electronics, their inherent compliance and freedom from electromagnetic interference. However, extracting meaningful information such as contact location and force magnitude from the light intensity patterns captured by such sensors is extremely complex, due to the non-linear and globally coupled nature of the data. In this study, we use a previously validated optical tomography-based sensor that uses concurrent illumination from 24 emitter optical fibres and receives light from 24 receiving optical fibres, embedded in a deformable transparent silicone layer. Using an RGB camera to capture deformation-induced light changes, we demonstrate that physics-informed delta feature representations, when combined with deep learning (DL) algorithms perform much better than the classic machine learning (ML) methods used in our previous study. Our proposed framework increase the localization accuracy to >97% along with R2value and Mean Squared Error (MSE) value for force estimation to >0.98 and2, respectively. Our results demonstrate that combining physics-informed feature engineering with deep learning enables robust extraction of tactile information from globally coupled, non-linear input data.