Bayesian State Estimation via Contact Sensing

We investigate the problem of using contact sensors to estimate the configuration of the environment during manipulation. Contact sensors are unique because they inherently discriminate between “contact” and “no-contact” configurations. As a result, the set of object configurations that activates a sensor constitutes a lower-dimensional contact manifold in the configuration space of the environment. This causes conventional state estimation methods, such as the particle filter, to perform poorly during periods of contact due to particle deprivation.

Our work introduces the manifold particle filter as a principled way of solving the state estimation problem when the state moves between multiple manifolds of different dimensionality. The manifold particle filter avoids particle starvation during contact by adaptively sampling particles that reside on the contact manifold from the dual proposal distribution. For low-dimensional problems, e.g. planar manipulation of a single movable object, we describe several techniques of sampling from the dual proposal distribution by constructing an approximate representation of the contact manifold.

Unfortunately, it is not tractable to construct an explicit representation of the contact manifold in high-dimensional problems, such as estimating the configuration of a manipulator. In this case, we implicitly represent the contact manifold as the iso-contour of a loss function and use constraint projection to sample from dual proposal distribution.

Our simulation results show that that the manifold particle filter outperforms the conventional particle filter in both speed and accuracy on two applications: (1) estimating the pose of a movable object relative to the hand during planar pushing and (2) estimating the configuration of a manipulator in a known environment. We additionally demonstrate the approach in real-robot experiments using feedback from tactile sensors on the BarrettHand and the iHY end-effector.