Bayesian online changepoint detection to improve transparency in human-machine interaction systems

This paper discusses a way to improve transparency in human-machine interaction systems when no force sensors are available for both the human and the machine. In most cases, position-error based control with fixed proportional-derivative (PD) controllers provides poor transparency. We resolve this issue by utilizing a gain switching method, switching them to be high or low values in response to estimated force changes at the slave environment. Since the slave-environment forces change abruptly in real time, it is difficult to set the precise value of the threshold for these gain switching decisions. Moreover, the threshold value has to be observed and tuned in advance to utilize the gain switching approach. Thus, we adopt Bayesian online changepoint detection to detect the abrupt slave environment change. This changepoint detection is based on the Bayes' theorem which is typically used in probability and statistics applications to generate the posterior distribution of unknown parameters given both data and prior distribution. We then show experimental results which demonstrate the Bayesian online changepoint detection has the ability to discriminate both free motion and hard contact. Additionally, we incorporate the online changepoint detection in our proposed gain switching controller and show the superiority of our proposed controller via experiment.