Simultaneous and proportional control of a prosthetic hand and wrist is still a controversial issue, although giant steps have lately been made in this direction. In this paper, we study the application of a novel machine learning method to the problem, with the aim to potentially improve such control. Namely we apply different kernels for tensor Gaussian process regression to data obtained from an advanced, flexible tactile sensor applied on the skin, recording muscle bulging in the forearm. The sensor is a modular, compact bracelet comprising 320 highly sensitive elements organized as a tactile array. The usage of kernel functions with tensor arguments and kernel distances computed on Riemannian manifolds enables us to account for the underlying structure and geometry of the tactile data. Regression accuracy results obtained on data previously collected using the bracelet demonstrate the effectiveness of the approach, especially when using Euclidean distance and Kullback-Leibler divergence-based kernels.