Tracking object pose and shape during robot manipulation based on tactile information

Despite the obvious relevance of tactile sensing to robot manipulation, no comprehensive theory exists regarding how to use this information to estimate hand-object state. This paper explores this issue in the context of particle filtering. The paper proposes using measurements of contact positions and velocities for the purposes of estimating the shape and pose of an object held by the robot. The likelihood of position and velocity measurements of manipulator contact surfaces touching the object and for those contact surfaces not touching the object are considered. Although calculating the measurement distribution involves a double integral over the object surface, this paper proposes a closed form solution for the special case of polyhedral objects that takes time linear in the number of facets. Quantitative results are reported that characterize the performance of the approach in terms of localization error and that characterize the relative contributions of different measurement types. In addition, the paper proposes a variation on particle filter annealing that is shown to accelerate localization without committing the system to a particular tracking solution. Experimental results demonstrates the performance improvement relative to solutions that do not use annealing.