A Gripping Idea
Combining deep learning with motion planning algorithms, researchers have reduced the time it takes to compute robot arm movements 300-fold.
Robotics promises to improve speed and accuracy — while reducing cost — in e-commerce warehouse picking. The importance of this has been put on display in the past year where, due to COVID-19, many people were relying on online purchases for essentials, yet capacity to fulfill those orders was reduced by the need for warehouse workers to distance themselves from one another.
Warehouse picking presents difficulties not seen in many other areas of automation. An assembly line, for example, requires that the same set of actions be performed again and again. Contrast that with a warehouse, where every order requires a customized motion plan to complete the task.
The problem is usually approached using a traditional motion planning algorithm, or deep learning. In the former case, motion is smooth and highly accurate, but the instructions to send the robot can take in excess of thirty seconds to compute. In the latter, instructions can be calculated very rapidly, but the motions are only approximate, and not suited to high precision applications. If only there were a solution that could provide the best of both worlds...
A team of researchers at the University of California, Berkeley have recently presented a hybrid grasp-optimized motion planning technique that promises to offer exactly that. They first use a deep neural network to give the process a warm start. This approximate control data is then fed into a motion planning algorithm as a starting point. Providing this head start allows the computation time to be reduced by two orders of magnitude. The end result is an accurate motion plan that can be calculated in milliseconds.
In addition to saving time, the neural network is also trained on data that is designed to limit jerks — rapid changes in acceleration — in the instructions passed along to the robot. This is particularly important with repetitive tasks like warehouse picking, where jerky motions can substantially reduce the lifespan of a robot.
The team tested their method on a physical UR5 robot fitted with a Robotiq 2F-85 parallel gripper. The robot was tasked with moving various objects from one fixed bin to another. They found that motion plans could be calculated in 80 milliseconds. By comparison, using a traditional motion planning algorithm, that computation time was 29 seconds.
The researchers believe that their work, along with other recent advancements in the field, will be transforming real world warehouse operations in the next few years.