Towards tactile intelligence for improving advanced robotic manipulation

Abstract

Current industrial needs—now characterized by the agile and lean paradigm—focus on producing smaller batches of greater variety. Since robots require more advanced skills to handle a greater variety of parts, efforts have been made to improve their manipulation capabilities. However, these efforts have mainly involved artificial vision, which is not enough to provide all the information needed for advanced manipulation tasks. And although non-vision methods exist, such as those based on tactile sensing, they are rarely used in industry. Perhaps industry would be more inclined to adopt tactile sensors if we could better exploit their signals. Thus, the main objective of this thesis is to advance the field of tactile intelligence, in order to improve robots’ manipulation capabilities. To do so, three methods are proposed. First, to improve robots’ ability to grasp fragile objects, we developed a dry gecko-inspired adhesive and used it to cover tactile sensors integrated on a gripper’s fingertips. We also patented a chevron pattern for the adhesive that compensates for tangential force and moments at low pressure. We showed that the interaction force between the grasped object and the adhesive is strongly correlated with the contact area—and unlike with typical Coulomb friction, it is less correlated with the amount of force. Therefore, this material is well-suited to handling delicate objects such as those that are fragile or deformable. Second, the problem of detecting and classifying important dynamic events using tactile measurements is investigated. Specifically, we applied sparse coding to the problem of detecting important dynamic events, such as an object slipping from a robot’s grasp. This method of encoding tactile sensing data reduces the amount of space required to store the data coming from the sensors, but it is also demonstrated that it highlights their most important features. In fact, we show that using only a low number of sparsely encoded features, it was possible to accurately detect four different classes of dynamic events. Third, we studied three ways to identify an object from among 50 non-fragile objects: measuring vibrations when the gripper moved across the surface of an object; measuring how wide the gripper opened during the grasp; and measuring object deformity on the tactile sensor. We found that all these methods are able to identify objects, but by far the most efficient is object deformity (i.e., perceived deformation of the object at the fingertips), which alone is enough to identify the vast majority of the objects. The benefits of this method are that there is almost no exploration phase and vision is not required, so objects can be identified at the same time as they are grasped—meaning there is low increase in processing time or operating cost.

Date	29 Apr 2020
Original language	American English
Awarding Institution	École de technologie supérieure
Supervisor	Vincent Duchaine (Supervisor) & Mark Cutkosky (Co-supervisor)