Nationwide we HAUL it ALL!  Services start at $9.95, ANY SIZE… 7 days a week year round.

Faster than Amazon, Hauling items within Hours!  Learn More about SERVICES

Haultail is Nationwide from Courier to Big and Bulky Rapid Delivery. Learn More about LOCATIONS

  • Download now!

Smarter Delivery Hinges on Smarter Robots


As customers expect to get packages faster, companies are turning to robots that can learn on their own

Companies and researchers world-wide are racing to develop artificial-intelligence systems that can enable warehouse robots to handle new and varied objects—telling the difference between a cardboard box and a small circuit board for instance—without the need for extensive additional training or human help.

Such capabilities would make machines more reliable in fulfillment centers, critical components of the e-commerce supply chain which at times is strained by a lack of a steady workforce and growing demands for speedier delivery.

In fulfillment centers, robots must handle millions of objects of different shapes, sizes and colors. Such variability can be a challenge for current machines, according to roboticists and warehouse operators. Some robot arms, for instance, use 3-D cameras to detect and grasp objects but are hampered by the bright glare of shiny products like packets of chips. Others struggle to grasp small objects like electrical components or soft, squishy things like certain fruits and vegetables.

Most automation technologies deployed today in warehouses aren’t intelligent, meaning humans have to program what a robot can do, according to Mehdi Miremadi, a partner at McKinsey & Co. who specializes in AI and robotics. New capabilities, like navigating a new path in a fulfillment center or grasping new products to package them for shipping, have to be hand coded, he said.

As supply chains have become more complex and consumers expect same-day delivery, that approach has become untenable, incentivizing warehouse operators to turn to systems that can learn on their own. In turn, startups are popping up to satisfy the demand.

U.S. venture-capital investment in logistics-focused robotics and automation companies since 2015 has topped $1 billion, according to data firm PitchBook.

“Warehouse robots are an exciting space again…There are many players trying to tackle this,” Mr. Miremadi said. “None of them are able to provide an end-to-end solution yet” that spans functions as varied as navigation planning and handling objects.

Often, the initial proof-of-concept application is grasping and sorting objects into shipping packages with robotic arms, where the cost of making mistakes is lower than in other instances, like autonomous navigation.

Peter Puchwein, vice president of innovation at Austrian warehouse logistics company KNAPP AG, said its robot pickers reliably handled only about 15% of objects before the company last fall deployed a new AI-for-robotics system developed by AI startup Covariant at a warehouse it operates for a German electrical supplies wholesaler.

The Berkeley, Calif.-based company is one of several commercial outfits trying to build flexible AI systems that work reliably in real-world settings, a challenge that has limited the use of robotics thus far to very structured environments, like factories.

KNAPP’s robot pickers outfitted with Covariant’s AI now reliably handle 95% of objects and they are slightly faster than humans, said Mr. Puchwein. The robots can pick about 600 objects an hour, compared with human workers who can pick between 400 and 450 an hour during an eight-hour shift.

The system, dubbed Covariant Brain, leverages a combination of machine-learning techniques popular among Silicon Valley’s biggest tech companies—including deep learning and reinforcement learning—which DeepMind, a division of Alphabet Inc.’s Google, used to build AlphaGo, the Go-playing software that defeated multiple world-class human players.

The technology enables robots to learn how to manipulate objects on the fly—without human help or hand-coding of the software—an important advance as it is impossible to fully anticipate the variety of objects a robot might encounter as it picks and packs merchandise across multiple facilities, for different clients, according to Covariant President Pieter Abbeel.

RightHand Robotics Inc., another robotics company leveraging AI, uses computer-vision software to help robots figure out how to pick up objects in the most efficient way. The company says its system’s accuracy is comparable to or better than human pickers.

KNAPP is planning to deploy the Covariant technology at a second facility in the U.S., according to Mr. Puchwein. No staff have been laid off because of the new technology, he added. Instead, they have been retrained to understand more about robotics and computers.

The 5% of objects the robots still struggle with are more a function of current gripping technology, he said. Robots sometimes still have a hard time with very soft objects.

“There’s so much more we have to do in distribution centers in order to fully automate them,” said Howie Choset, a robotics professor at Carnegie Mellon University. “Companies have only begun to scratch the surface.”


This article was originally published on

We updated our privacy policy as of February 24, 2020. Learn about our personal information collection practices here.