As the use of online shopping has skyrocketed, so have consumers’ expectations for quick and accurate deliveries. This has led to a pace of activity in warehouses that is incredibly demanding on workers.
Assistance from robots could help ease the workflow and optimize output, but it’s been a challenge to integrate them into these types of workspaces. In assembly lines, robots perform the same tasks repetitively. But in warehouses, robots need to grasp and move objects of varying shapes and sizes, requiring them to make many adjustments. If their movements are jerky, it can damage the object, as well as lead to wear and tear on the robot.
But a solution may be at hand. A research team led by Ken Goldberg, professor of industrial engineering and operations research and of electrical engineering and computer sciences, has developed artificial intelligence software that allows robotic arms to quickly grasp and move objects in a smooth and steady motion.
Previously, Goldberg and postdoctoral researcher Jeffrey Ichnowski had created a grasp-optimized motion planner that computed the most efficient way for robots to handle and move objects, but when they modified the software to smooth out the motions, it significantly slowed everything down, making the system impractical.
To address this challenge, the research team — including graduate student Yahav Avigal and undergraduate student Vishal Satish — combined a deep learning network with the motion planner. After sampling thousands of motions that a robot would likely make, they used them to train the deep learning network, which then proposed trajectories of varying lengths and selected the optimal one. The motion planner used that information to identify the best starting point for planning the smoothest trajectory to use. This integration proved to be key, allowing the team to cut the average computation time from 29 seconds to 80 milliseconds, or less than one-tenth of a second.
The researchers hope to expand this work to other robotic tasks in complex environments, as well as test new approaches to further speed up the deep learning process.
Learn more: Deep learning can accelerate grasp-optimized motion planning (Science Robotics)