Local implicit surface estimation for haptic exploration

Abstract

Autonomous grasping and manipulation of unknown objects is a central skill for humanoid robots. This is particularly challenging, as shape information needs to be obtained from sensory data which is often noisy and incomplete. However, object shape information is usually a key prerequisite for grasp and manipulation planning and thus needs to be estimated even if the available sensor data is limited. We propose a method for implicit surface modeling based on sparse contact information, as it arises e.g. from haptic exploration. Surfaces are locally defined using the contact points and their normals, and the object shape is extrapolated by integrating this partial information. For each contact contributing to the estimation, the local convexity or concavity is determined depending on its neighbors and their respective normals. Taking into account contact positions, normals and local convexities or concavities, the Implicit Shape Potential of the overall surface is generated. In contrast to popular methods based on Gaussian Processes, this representation allows for local details like edges and corners, without losing the ability to interpolate in the case of noise. In addition, it provides information to guide iterative exploration algorithms. The proposed method is evaluated on a set of various 3D shapes that possess flat and curved surface regions as well as convex and concave edges.