A UR-5 robot performs a gripping movement
Adam Lau / Berkeley Engineering
Pick-and-place robotic arms for packing boxes in warehouses can now work more than 350 times faster because a neural network predicts how quickly they can safely move items.
The coronavirus pandemic has led to an increase in online shopping. “Vendors have a very hard time meeting demand,” says Ken Goldberg of the University of California at Berkeley.
Goldberg’s lab previously developed software that improves the grasp of a type of robot often used in warehouses. Computer vision can be used to determine where in three-dimensional space an object to be gripped is located relative to the robot’s claw.
“Now the bottleneck has passed to the movement side of things when the object is within reach,” says Jeffrey Ichnowski, also in Berkeley.
Robots can move quickly, but not always safely. The amount of “jerk” or rapid change in acceleration can mean the difference between a successfully delivered package and one thrown on the floor. A jerk can also lead to wear on the robot arm and shorten its service life.
“We have to stand on the edge of the limits of speed, acceleration and jolt,” says Ichnowski. However, it takes time for robots to find the safest but fastest way to make a difference.
Ichnowski and Goldberg and their colleagues have added what is known as a neural network to their robot software. The network then had an assessment of how the robotic arm behaves when it moves thousands of objects over several weeks.
Eventually, the neural network learned to determine the best motion path for a given scenario within 80 milliseconds. The existing software took 29 seconds to perform the calculation.
“Gradually changing operational response times as outlined in this document will make a huge difference for warehouse operators,” said Andrew Lahy of Cardiff University, UK, and director of solution design at logistics company DSV.
“We think this is practical and can be used relatively quickly,” says Goldberg.
Journal reference: Science Robotics, DOI: 10.1126 / scirobotics.abd7710
More on these topics: