In 2017, the University of California, Berkeley (UC Berkeley) and OpenAI created Embodied Intelligence to address one of the challenges of robotics: teaching robots what to pick and how to grasp it outside of structured parameters. Two years later, the company has renamed itself Covariant and has seen its system reliably work in a German warehouse for four months.
“From the very beginning, our vision was to ultimately work on very general robotic manipulation tasks,” said Pieter Abbeel, cofounder of Covariant. “The way automation’s going to expand is going to be robots that are capable of seeing what’s around them, adapting to what’s around them and learning things on the fly.”
The company spent nearly a year speaking with various companies in different sectors about how smarter robots could impact their businesses. It quickly became clear that manufacturing and logistics were in high demand, especially enhanced automation for tasks that typically require human workers such as picking.
The main hurdle has been the limitations of robotic pickers to pick only specific parts or that shifting to robotics requires a costly investment and that it is nearly impossible to train the robots on every individual part at a plant. Covariant’s solution uses simple hardware—including a standard industrial arm, suction gripper and 2D camera system—but combines it with a massive neural network.
“We can’t have specialized networks,” Abbeel explained. “It has to be a single network able to handle any kind of SKU, any kind of picking station. In terms of being able to understand what’s happening and what’s the right thing to do, that’s all unified. We call it Covariant Brain, and it’s obviously not a human brain, but it’s the same notion that a single neural network can do it all.”
The key component of Covariant’s success has been its approach in using artificial intelligence (AI). Unlike traditional methods where robots are taught a specific task, Covariant’s robots learn general abilities. The robots’ ability to master general skills, such as 3D perception, real-time motion planning and an object’s physical affordances, makes it easier for them to adapt to different tasks by breaking them down to determine the steps required to complete the task.
“Our system generalizes to items it’s never seen before. Being able to look at a scene and understand how to interact with individual items in a tote, including items it’s never seen before—humans can do this, and that’s essentially generalized intelligence,” Abbeel said. “This generalized understanding of what’s in a bin is really key to success. That’s the difference between a traditional system where you would catalog everything ahead of time and try to recognize everything in the catalog, versus fast-moving warehouses where you have many SKUs and they’re always changing.”
Like with any new technology, success and failure rates are vital. In the four months the system has been in use at the German plant, the company said its robot has gone from being able to pick around 15 percent to 95 percent of its product range. The robot has learned to pick and sort more than 10,000 items with more than 99 percent accuracy.
Interested in more robotic innovations in manufacturing? Check out Tackling Robotic Bin-Picking Challenges with Motion Control and Path Planning and Robot Arm Uses AI to Get a Better Grip.