Learning a tool's homogeneous transformation by tactile-based interaction

We propose a tactile-based manipulation strategy to learn the homogeneous transformation of a grasped rigid tool, using tactile sensing delivered through a tactile matrix sensor covering the tool surface. Exploiting the self-learning tactile servoing controller, a robot safely use the tactile tool to implement different tactile-based exploration primitives (EPs). Considering EPs as input and observing the tactile contacts as output, the robot can robustly estimate the tool's homogeneous transformation. The learned transformation are combined with the known robot's kinematics model to form a new tool manipulation kinematics chain, thereby realizing a step towards a “plastic body schema” for flexible tool use by a robot. We numerically evaluate the method's feasibility and robustness assuming that measurements are only polluted by Gaussian white noise, then evaluate the proposed method with a real robot setup — a KUKA LWR and a SCHUNK SDH-2 hand grasping a rigid tactile tool. With the new manipulation chain, we demonstrate two tactile tool's servoing experiments: reactively sliding and rolling tool and tracking an unknown object edge.