Gelsight robotic fingertip

Abstract

Tactile sensing is integral in robotic manipulation, enabling robot hands to identify objects and explore the environment. The gelsight tactile sensor designed in MIT CSAIL is able to reconstruct the surface topography of objects up to submicron accuracy. This thesis is focused on the integration of the gelsight tactile sensor into the MIT Team's gripper for the 2017 Amazon Robotics Challenge. The design of the sensor is dependent on the three manipulation primitives in the gripper platform: scooping, grasping, and suction. This paper discusses the primitives and the sensory inputs they would require to create more precise manipulation models. We then propose a gelsight tactile sensor around those primitives, creating a design that can recover shear, deflection, compliance, and contact area information. Lastly, we implement some of the image processing to recover deflection by creating a mapping between image points and ground truth vicon values, producing an 75% accuracy rate within a distance threshold, a value that can be improved with increased measurements and better image processing techniques. Further work focuses on image processing for shear, compliance, and contact area, to be integrated into the overall design for the 2017 ARC gripper.