Contact state estimation using machine learning

In this paper we present an approach that uses machine learning to determine the location of a contact between a gripper and a T-bar valve based on force/torque sensor data. The robot performs an exploratory behaviour that produces distinct force/torque data for each contact location of interest: no contact, a contact aligned with the central axis of the valve, and an off-center contact. Probabilistic clustering is utilised to transform the multidimensional data into a one-dimensional sequence of symbols, which is then used to train a hidden Markov model classifier. We present the results of an experiment where the learned classifier can predict a contact location with an accuracy of 97% on an unseen dataset.