CEO DailyCFO DailyBroadsheetData SheetTerm Sheet

The robots that pack bread during the pandemic

July 21, 2020, 3:11 PM UTC

During the pandemic, robots are the literal breadwinners.

As Samir Menon, the founder and CEO of robotics startup Dexterity, told Fortune, the “coronavirus took us a bit on a wild ride.” Dexterity, a roughly two-and-a-half-year-old startup that’s raised $56.2 million in funding from investors like Kleiner Perkins and Lightspeed Venture Partners, specializes in software that makes industrial robots, like mechanical grippers and machines that pick-and-place items, more capable.

Among the customers that have relied on Dexterity’s technology during the pandemic is Bimbo Bakeries, which produces some well-known baked goods via brands like Sara Lee, Oroweat, and Boboli. Because Dexterity is working with Bimbo Bakeries, the startup has been deemed an “essential business,” since the food supply chain needs to continue operating unperturbed amid the lockdowns.

In an online demonstration from its Redwood City, Calif. headquarters last week, Menon showed Fortune an example of how robots outfitted with the company’s technology help prep pre-packaged loaves of bread for delivery to places like grocery outlets.

For the demonstration, two robots with mechanical arms cruised down a railing that ran between two sets of blue crates, each loaded with various loaves of bread. The robots needed to coordinate with each other so they didn’t bump into one another while they transported loaves from one set of crates to the other—a common task used for facilitating bread shipments.

Dexterity demonstrates how its software helps robots pick-and-place loaves of bread.
Dexterity demonstrates how its software helps robots pick-and-place loaves of bread.

Using computer vision technologies, a robot could recognize when one empty crate on its right side lacked enough bread loaves, triggering its mechanical arm to pick up a bread loaf from a crate on its left side. The robotic arms can figure out how much pressure is needed to apply to a particular object, which is helpful so they don’t smash a spongy material like bread when they grip it. The software also helps the robot decide to gently lower the bread and other objects into a crate instead of dropping it carelessly into a container, which Menon described as a sort of Achilles heel for some pick-and-place machines. 

Although Dexterity’s software helps the robots map out their surroundings so they can maneuver autonomously throughout a facility, Menon said the robots are typically affixed to a railing so that people can rest assured the machines won’t suddenly go somewhere they aren’t supposed to.   

“It’s important for them to feel comfortable,” Menon said regarding humans who work alongside the robots.

Menon said that companies are currently using Dexterity’s technology at a few unspecified facilities on the “east and west coasts” of the U.S. He declined to comment on the number of robots that have been outfitted with Dexterity’s software at the Bimbo Bakeries facilities nor other facilities.

Dexterity is a young company, so it’s still refining its technology for customers. During one demonstration for Fortune, a robotic arm installed with the startup’s software dropped a glass measurement cup onto a table instead of gently placing it down, and while the beaker didn’t break, it was a definite no-no. Menon noted prior to that particular demonstration that translucent objects like glass containers are a “bane of any computer-vision” technology. 

Still, as long as the pandemic rages on, it’s likely companies with big supply chains like Bimbo Bakeries will continue turning to companies specializing in robotics in order to run their deliveries without too many hiccups.

“We feel grateful to be actually shipping bread,” Menon said.

Jonathan Vanian 


Deepfakes strike again. Outlets like the Jerusalem Post and the Times of Israel published editorials from a person named Oliver Taylor, who Reuters reported appears to be a deepfake and not an actual human. It’s unknown who created Taylor, but the report said his stories “reveal an active interest in anti-Semitism and Jewish affairs.” Unlike other deepfakes, the manipulated photo of Taylor doesn’t show any obvious distortions, like an elongated earlobe, that may tip people off that it’s a fake. But Reuters interviewed experts who noticed some flaws that led them to believe the photo was a deepfake. “The distortion and inconsistencies in the background are a tell-tale sign of a synthesized image, as are a few glitches around his neck and collar,” digital image forensics pioneer Hany Farid told the publication.

South Korea’s ‘Untact’ plans. South Korea is pushing heavily into A.I. as part of its “untact” initiative, which Bloomberg News described as “a future where people increasingly interact online and companies replace humans with machines to immunize themselves against the effects of rising wages and a rapidly aging workforce.” The advent of COVID-19 has amplified the country’s efforts to beef up on technologies like robots and self-driving cars that could “reduce the need for person-to-person contact,” the report said. The initiative is also intended to “reduce South Korea’s reliance on China as an export market” and to help “raise an army of 100,000 specialists in AI.”

Where’s the consent? Multiple federal lawsuits were filed against Amazon, Alphabet, and Microsoft for allegations that the tech giants trained their facial-recognition technologies with datasets that contained photos of people who never consented to have their visages used for those purposes, among other issues, CNET reported. “The lawsuits, filed in California and Washington state courts where the companies are based, seek class-action status, as well as monetary damages and restraint of defendants' activities pertaining to the database,” the report said.

Possible hope in a massive dataset. Over 100,000 people have registered to take part in a project by the U.S. Department of Health and Human Services to accumulate more participants for clinical trials related to potential COVID-19 treatments and vaccines, Oracle, who is helping with the effort, said. The organizers behind the project, known as the COVID-19 Prevention Trials Network, hope to eventually have millions of volunteers.


Accenture has chosen Sanjeev Vohra to be the consulting firm’s global lead of applied intelligence, replacing Athina Kanioura, who is leaving to another unspecified firm. Vohra previously oversaw business strategy and investments for Accenture as its growth and strategy lead for technology.

The University of California, Santa Cruz has picked Adwait Ratnaparkhi to be the executive director of the institution’s new master’s degree program in natural language processing (NLP). Ratnaparkhi was previously the director of voice and NLP research at Roku and a senior director of research at Yahoo.


Image recognition fail? Researchers at the Massachusetts Institute of Technology published a paper about the popular ImageNet dataset and possible problems during the data collection and curation phase that may impact the quality of image-recognition technologies trained from the giant corpus of photos. One of the problems: Photos within ImageNet that contained multiple objects were often labeled to describe only one object by the human annotators, which could “lead to the selection of images with multiple valid labels or even to images where the dataset label does not correspond to the most prominent object in the image.”

As VentureBeat described in a report about the research, the authors “conclude that it can be difficult for human annotators who are not experts to accurately label images in some instances. Choosing from one of 24 breeds of terriers could be difficult unless you’re a dog expert, for example.”


A.I. can help solve America’s education crisis—By YJ Jang

Startup Graphcore takes on Nvidia with latest A.I. chip—By Jeremy Kahn 

How chatbots are helping in the fight against COVID-19—By Jonathan Vanian

Meet Dubsmash, the video service stealing some of TikTok’s thunder—By Danielle Abril


Deep learning don’t come cheap. In order for deep learning to be more widely adopted, the computing costs associated with the software needs to drop. That’s one takeaway in a Wired article about new research describing how progress in A.I. partly depends on an ever-increasing amount of computing power.

From the article:

AI’s appetite for computation has risen remarkably over the past decade. In 2012, at the beginning of the deep-learning boom, a team at the University of Toronto created a breakthrough image-recognition algorithm using two GPUs (a specialized kind of computer chip) over five days. Fast-forward to 2019, and it took six days and roughly 1,000 special chips (each many times more powerful than the earlier GPUs) for researchers at Google and Carnegie Mellon to develop a more modern image-recognition algorithm. A translation algorithm, developed last year by a team at Google, required the rough equivalent of 12,000 specialized chips running for a week. By some estimates, it would cost up to $3 million to rent this much computer power through the cloud.


Update July 21, 2020: A Dexterity spokesperson clarified that that the startup's customers are using its technology at unspecified locations on the East or West coast, not specifically Bimbo Bakeries.