The essential components are a large neural network, a 2-D camera, and suction cups
Covariant’s AI-powered robotic arm relies on reinforcement learning, as its neural network trains itself to find items through trial and error. Covariant CEO Peter Chen said the neural network started off understanding how to sort a set of 10,000 items with a digital simulation – a bin full of random items. Later, Chen and the other Covariant founders transferred the system to a robot equipped with an industrial arm, a 2-D camera system, and three suction grippers. The robot’s first gig is at Obeta, a German electrical supply wholesale company outside Berlin.
As the robot picks out objects that are upside down and sideways, the results stream back to the “Covariant Brain,” the neural network for all the company’s robots. Covariant provides that the general abilities that the robot learns include few-shot learning and real-time motion planning. Covariant’s co-founders perfected their design while engaged in AI research at UC Berkeley and OpenAI, an independent research lab based in San Francisco. Their robot joined Obeta’s team in late 2019, an effort that was the culmination of two and a half years of development and testing.
The sorting robot, which currently works alongside human sorters, does its job with over 99 percent accuracy. Its presence at Obeta, a company that has been open since 1901, is a collaboration between Covariant and Knapp, an Austrian warehouse logistics tech company. Knapp plans to distribute “Covariant-enabled” robots to customer warehouses such as Amazon’s in the next few years. Yet Covariant has a larger aim, to bring its technology to a wide spectrum of other industries, including electronics manufacturing, car manufacturing, textiles, biology labs, construction, farming, hotels, and even elder care.
Covariant’s robot, which sits atop a small platform just above a conveyor belt, takes up far less room than a person and doesn’t take breaks. This will allow retailers to solve a labor shortage, minimize use of space, and maximize efficiency all at once. The potential for solutions extends beyond picking to induction, the choosing of an object from a mix and placement of it onto a small compartment of a moving conveyor belt, and putwall, the choosing of an object from a mix and placement of it into a static shelving system.