Raphael Thys
FedEx unveiled a two-armed robot called DexR this week that’s designed to automate one of the trickiest tasks facing the company’s human employees—loading a truck with packages.
The new robot aims to use artificial intelligence to stack rows of differently sized boxes inside a delivery truck as efficiently as possible, attempting to maximize how many will fit.
That task is far from easy for a machine. “Packages come in different sizes, shapes, weights, and packaging materials, and they come randomized,” says Rebecca Yeung, vice president of operations and advanced technology at FedEx. The robot uses cameras and lidar sensors to perceive the packages and must then plan how to configure the available boxes to make a neat wall, place them snugly without crushing anything, and react appropriately if any packages slip.
Pause
Fedex’s DexR robot stacks packages inside of a delivery truck (part of the video is sped up).
“A few years ago, AI was not at a stage where it was smart enough to handle this kind of complex decision-making,” Yeung says. DexR is currently in testing, ahead of a wider rollout at FedEx at some point in the future.
While generative AI tools like ChatGPT have created a sense in many industries that AI technology is ready to take on just about anything, handling objects in the messy, unpredictable real world still poses formidable challenges for algorithms. Most industrial robots are designed to carry out highly repetitive jobs with extreme precision, but no variation.
Roboticists are making progress. A growing number of machines now use AI to do things like recognize objects or determine how to grasp them. This can involve training algorithms inside a simulation where errors matter little before transferring that software to a real robot. But making the jump from simulation to the real world is notoriously difficult.
Better algorithms, new approaches to using machine learning for robots, and improved hardware and sensors have all started to open more commercial applications for advanced robots.
“In the last year or two, people have taken advances in AI and machine learning and said ‘we can make a real business case here, whether it’s lowering costs or improving efficiency or whatever,” says Matthew Johnson-Roberson, director of the robotics institute at Carnegie Mellon University.
Johnson-Roberson says years of investment in areas like self-driving vehicles, combined with a steady cadence of advances in AI, will allow robots to creep into more workplaces. “My hope is that we’re just at the beginning of a coming wave in commercial robotics.”
Pause
Dexterity’s AI software uses data from cameras and lidar sensors to perceive packages and plan how to stack them.
The FedEx robot was built for the company by Dexterity, a startup based in Redwood City, California, that specializes in developing robotic systems for various warehouse tasks, using AI.
Dexterity CEO Samir Menon says the robot built for FedEx uses generative AI to work out how to stack boxes of various kinds. It also uses AI to identify and grab the boxes. But these systems need to be woven together with careful engineering, Menon says.
Each time it places a box in a stack, the system uses force-feedback to ensure the package fits tightly, and it also scans the stack using cameras and depth sensors to see how it compares to its existing model. Any discrepancy will require the robot to adapt its stacking plan as it goes along.
The growth of ecommerce—and particularly Amazon—has turned working with packages into an innovative frontier for robot development. Amazon is currently rolling out thousands of more advanced robots as it continues to squeeze greater efficiency out of the facilities where it stores and processes products.
Truck unloading and loading “presents a harder challenge” than the picking work that robots currently do in warehouses because it happens in a tight space with a variety of boxes, says Pulkit Agrawal, a professor at MIT who specializes in AI and robotics. There are ways to engineer the system that can “reduce the complexity,” but the demo is still impressive, he says.
If AI accelerates the adoption of robotics, it might spark fears of job displacement. The ongoing US auto worker strike is partly related to technological trends sweeping that industry, including electrification and autonomous driving.
Yeung says the robot is still being perfected, but it should eventually load a truck as quickly as a skilled human. FedEx already uses robot technology developed by another company, Berkshire Grey, to sort parcels inside some facilities. It spent $200 million on these systems in 2022.
Yeung declined to say how many of the robots FedEx will deploy or how quickly, and data on their reliability is still being gathered. But the capabilities demonstrated by Dexterity’s robot should transfer to other tasks so that robots can take on more work at FedEx. “This is a big deal for us,” she says. “We’re excited about these next-gen capabilities improving our operations.”