Robots have learned to grab and move objects with ease

In 2020, restrictions due to the pandemic have made online shopping more popular than ever. However, the rapidly growing demand leads to delays in shipment of goods and dangerous situations in the warehouse. Scientists have created new AI software that improves the skills of robots to quickly grip and move objects smoothly. Soon it will help workers in the warehouse.

Automating warehouse tasks can be challenging. Many actions that are natural for humans are actually quite challenging for robots. For example, deciding where and how to get different types of objects. It is also important to further coordinate the movements of the shoulders, arms, and wrists, which are necessary to move each of them from one place to another. In addition, the movement of robots is sharper – this increases the risk of damage to both objects and robots.

“Warehouses are still mostly manned because it is still very difficult for robots to reliably grip many different objects,” explains Ken Goldberg, senior author of the study. – On an automotive assembly line, the same movement is repeated over and over, so it can be automated. But in the warehouse, all orders are different.”

In earlier work, Goldberg and UC Berkeley researcher Jeffrey Ichnowski created a grasp-optimized motion planner. However, the engineers failed to make them smooth. Although the software parameters were adjusted to create smoother movements, these calculations took an average of about half an hour.

In a new study, Goldberg and Ichnowski collaborated with UC Berkeley graduate student Yahav Avigal and student Vishal Satish to dramatically speed up the computation time of the motion planner by integrating a neural network with deep learning.

Neural networks allow a robot to learn by example. Later, the robot can often generalize similar objects and movements. However, these approximations are not always accurate enough. Goldberg and Ichnowski found that the approximation generated by the neural network can then be optimized using a dedicated scheduler. By combining a neural network with a motion planner, the team reduced the average computation time from 29 seconds to 80 milliseconds.

The study authors are confident that thanks to this and other advances, robots will soon be able to help warehouse workers. “This is a great new opportunity for robots to support humans,” concludes Goldberg.