NRI: FND: Scalable Multimodal Tactile Sensing for Robotic Manipulators in Manufacturing

The research focuses on sensors and software to enable robots to have an artificial sense of touch. Current robot hands and grippers rarely have the ability to sense their environments; they operate by precisely repeating pre-programmed motions. This prevents the use of robots in applications where the state of the world is unknowable in advance, since such use requires a robot to sense and react to unexpected events. The project will design novel and sophisticated touch sensors and equip robot hands with them. The sensors will include multiple types of touch sensors and integrate them to detect different types and characteristics of contacts: fast versus slow, transitory versus maintained, etc. These data will be interpreted by software algorithms that learn how to make use of the newly acquired touch data. This approach takes its inspiration from the human hand, which is also equipped with multiple types of tactile sensing elements relaying information to the nervous system. The proximate goal of the work is to build tools for material handling in cluttered environments, a task that can enable e-commerce and supply chains to run more efficiently, help manufacturers become more efficient and competitive, and reduce injuries to workers by assisting with repetitive tasks that are known to cause injuries.

Novel sensor design, computation and planning will be integrated to endow robots with a sense of touch, thereby enabling manipulation in cluttered environments. At the hardware level, the sensors will combine piezoelectric and resistive strain sensing to capture both transient, high-frequency responses and absolute strain measurements. Sensor sheets will be stacked in multiple sensing layers to provide a rich signal set that departs from the 'one location, one taxel' approach that is normally taken and maximum use will be made of the rich data by combining model-based and data-driven approaches to define motor control primitives. These will be combined to implement algorithms for motor control and planning that use the tactile data in a target task: bin-picking and kitting in manufacturing. The overall goal is a compact, multi-modal, scalable tactile sensing system for robotic manipulators, along with the low-level motor skills and high level planning algorithms that use tactile data for manipulating in clutter.