ABSTRACT
Daily, millions of object parts are assembled by hand, even when objects are rigid and well-known. A natural question follows: what is missing in robotics in order to automate these tasks? While in recent years picking solutions have made significant progress in warehouse automation, dexterous robotic manipulation often requires more than just a good grasp. In this talk, we will discuss how robots must also learn to accurately observe and manipulate the pose of grasped objects to achieve more complex manipulations. I will exemplify this in the case of kitting, an omnipresent task in assembly pipelines that consists of grasping a target object from an unsorted set of objects and placing it accurately in pre-established configuration.
Kitting, like most manipulation tasks, requires a good perception system that can infer the state of manipulated objects. Because contact is unavoidable, tactile sensing becomes a rich source of state information. In recent work, we have shown that even tactile alone can be used to accurately localize manipulated objects. To put this results in practice, I will also present our first steps of developing a self-supervised kitting system that can achieve high performance, while allowing the flexibility to adapt to novel scenes.