Modeling a Tactile Skin for Learning Fine Manipulation with a Robotic Hand

Kurzfassung

The dextrous manipulation of small objects is a key skill required for all kinds of industrial tasks like sorting or assembling. Humans excel in fine manipulation using their dextrous hands in combination with their spatially resolved sense of touch. In robotics, only recently, the advent of modern learning methods based on excessive training in simulation enabled breakthroughs in dextrous manipulation with multi-fingered hands, e.g., in-hand manipulation. However, for learning fine manipulation tasks, in addition a precise but computationally efficient simulation of tactile sensors is needed. Here, we present a novel model of a tactile skin that can be used together with rigid-body (hence fast) physics simulators. Tactile skins are especially attractive as they can be easily applied on the robot structure in contrast to the more commonly used bulky camera-based sensors. The model considers the softness of the real fingertips such that a contact can spread across multiple taxels of the sensor depending on the contact geometry. We calibrate the model parameters to allow an accurate simulation of the real-world sensor. For this, we present a self-contained calibration method that relies solely on the proprioceptive sensors of the fingers. A cap with a spherical indenter is attached to one finger, which presses on the sensor attached to another finger. To demonstrate the validity of our approach, we learn two challenging fine manipulation tasks in simulation: Rolling a marble and a bolt between two fingers. We show that, especially for the marble, tactile feedback is crucial for precise manipulation. Moreover, we demonstrate that all policies can be successfully transferred from the simulation to the real robotic hand.