Contact State Estimation Using Multiple Model Estimation and Hidden Markov Models

Abstract

In this paper we present an approach to estimating the contact state between a robot and its environment during task execution. Contact states are modeled by constraint equations parametrized by timedependent sensor data and time-independent object properties. At each sampling time, multiple model estimation is used to assess the most likely contact state. The assessment is performed by a hidden Markov model, which combines a measure of how well each set of constraint equations fits the sensor data with the probability of specific contact state transitions. The latter is embodied in a task-based contact state network. The approach is illustrated for a three-dimensional peg-in-hole insertion using a tabletop manipulator robot. Using only position sensing, the contact state sequence is successfully estimated without knowledge of nominal property values. Property estimates are obtained for the peg dimensions as well as the hole position and orientation.