This project will improve the ability to control fabric-reinforced inflatable soft robots. Soft robots offer tremendous promise because they can conform to features in an unknown environment and then change shape to perform tasks, such as entwining objects or squeezing into small gaps. These could be useful in search and rescue, disaster relief, and increasing customization in manufacturing. While scientists and engineers have managed to demonstrate these types of movements by trial and error, getting them to work under computer control has been difficult. Because soft robots are usually made of elastomeric materials, a multitude of sensors can be molded into the robot, measuring touch, proximity, and the shape of objects in the environment. With such a large number of sensors, the robot has access to a rich amount of data that can be used to help determine the current position and shape of the robot. The elastomeric material of soft robots also offers the opportunity to incorporate active materials that can change their properties upon command from a computer. These materials can be used to tune the stiffness of the soft robot or provide additional options for changing its shape. If successful, this research project will produce improved algorithms to process these sensor inputs so that the soft robot can distill them into meaningful information about itself and the world. The techniques will be leveraged to broadly spread the benefits of the research to the wider community. Teams of selected K-12 teachers will participate in a decathlon-style contest, using a single robot to compete in a series of diverse tasks and building their skills with soft robot technology for later use in their classrooms.
The research effort centers around developing concise models for the motions of a fabric-reinforced rubber tube that can be evaluated quickly by a computer. Particular attention will be given to behaviors that are difficult to represent by existing techniques, such as wrinkling, pleating and buckling. Based on the simplified model, the algorithm will postulate a variety of possible shapes for the robot at each instant. The algorithm will then decide which is correct by comparing measurements from its many sensors to the predictions of the model. Using the data gathered by the robot platforms, the robot will be able to learn how to select the appropriate commands -- that is, the combinations of pneumatic valve signals and tunable stiffness patches in the rubber walls -- to autonomously complete useful tasks where the robot must push on the environment. The algorithm will be validated in representative tasks such as hanging drywall and retrieving objects in clutter.
This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.
