People with disabilities, the infirm, and the elderly would benefit from capable service robots in the home that could assist with daily living tasks. Robots with arms and hands that have human-like abilities could help retrieve dropped or distant objects, and help with cooking, cleaning, and grooming. This would increase people's quality of life and independence, and greatly reduce the burden on their caregivers. While robots are found in almost every factory today, they live behind safety barriers, out of reach of people. Traditional robots are good at controlling their motion, but bad at controlling the forces they apply when touching objects; essentially, they are precise but not gentle. This research introduces a new way to design robots that allows them to easily feel when they are in contact with people and objects. A new design allows all of the motors that move the arms to be located in the robot's base. A traditional robot has a motor mounted at each joint in the arm, making the robot very heavy. When robots are lighter and move more gently, it does not just make them safer, but it also makes it easier to program them to learn how to pick up and manipulate objects. Traditionally, robots depend only on cameras to detect objects and plan how to grab them, but in this research, robots will also adapt and learn to interact with objects through their sense of touch.
Reliable and safe manipulation is a prerequisite for the vision of autonomous and semi-autonomous assistive robots, particularly towards the goal of increasing the length and quality of time that people can live independently. The key challenge is developing manipulator hardware that can move accurately but gently and supports safe contact-rich physical interaction with people and delicate objects. In this work a remote-direct-drive (RDD) configuration is used: all motors are mounted in the base, remotely connected to joints in the arm and hand via rolling-diaphragm sealed hydrostatic transmissions, which keeps the arm moving mass very small. Large radius direct-drive motors are used to achieve minimal joint static friction and allow joint compliance to be modulated over a wide range. Thus, the arm exhibits smooth second order joint dynamics, making it easier to plan contact-rich trajectories using optimization methods or learn contact-rich policies using off-line deep reinforcement learning. Moreover, the controllable compliance will enable our manipulation policies to use compliance when needed to accommodate environmental uncertainty but high stiffness when high precision is required. This type of manipulator has the potential to be lighter, lower friction, and more inherently backdrivable than traditional robotic arms, providing increased safety for human-robot interaction, and increasing the range of physical human assistance tasks that can be undertaken by autonomous systems.
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.
