In a newly renovated office building in Watertown, Massachusetts, 100 robots are slowly learning to pick up random objects, including everything from bottles of hand lotion to packets of Welch’s Fruit snacks.
Each human-like machine is equipped with four cameras: one on its head, one on its chest and two on its claws — and fixed to a stationary box. A few employees wearing smart devices inspect the work of the machines to see if they are completing their tasks correctly.
At this point, most of them don’t.
The robots, named Sonny after “I, Robot” character in the movie, have only been operational for a few days, and it will take several months of research and development before Tutor Intelligence has collected enough data for industrial applications, CEO Josh Gruenstein said during a tour of the facility on April 22. One day Sonny could be deployed to the operations of a Fortune 500 company.
Tutor has just moved to its new headquarters in Watertown, Massachusetts, located at the former home of Boston Scientificalong the Charles River. The historic site was previously a cotton mill in the 1800s.
“This is really a discovery tool for us to determine what the best methods are for learning robots and to start building that set of capabilities where we feel comfortable going to a customer,” he said of the robot fleet.
DF1 and the unique way Tutor teaches its robots
Data Factory 1, or DF1, is what Tutor calls “the largest data factory in the United States.” The fleet of robots is autonomously learning to move and manipulate objects with its claws, powered by the first startup. vision-language-action model, Ti0and through large-scale human oversight. In addition to the on-site team, remote workers in the US, Mexico and the Philippines monitor and correct the robot’s behaviors.
Tutor, founded in 2021, was started by Gruenstein and Alon Kosowsky-Sachs, Chief Technology Officer, MIT Laboratory for Computing and Artificial Intelligence. As students, they developed a keen interest in the robot brain and later set out to discover a “scalable recipe” to teach robots how to perform general tasks across their company.
The startup has gone up a total of $42 million to scale its robotic AI workers, including a $34 million Series A funding round that closed in December.
Instead of using the latest wearable or sensor technologies, Tutor chose their recipe—a mix of cameras, vision-learning software, and human supervisors—because it was more cost-effective.
“Actuators are expensive, sensors are expensive, even the physical structure of the robots … they’re not that big and could be 10 times cheaper,” Kosowsky-Sachs said. The cameras are cheap, in part, because of the iPhone and its large scale of production, he said. He and Gruenstein hope that economies of scale will eventually prevail for other parts used in their robots.
“There’s this fundamental problem with robotics, which is that you can build the smartest robot and the most capable robot in the world, but if [return on investment] it’s not extremely competitive, your business is not investable, and you’re probably going to go out of business,” Kosowsky-Sachs said.
A Tutor Intelligence employee walks through a row of robots, inspecting the machines to make sure they are working properly.
Nathan Owens/Manufacturing Dive
Support for AWS and its cloud services
Last year, Tutor participated in the inaugural cohort of the Physical AI Scholarshipan eight-week startup accelerator supported by Amazon Web Services, Nvidia, and MassRobotics, a nonprofit robotics center located in Boston’s Seaport District.
As a member, Tutor received dedicated science and engineering support from AWS and Nvidia, as well as $200,000 in AWS credits to help them “get their solution and, of course, to market,” said Alla Simoneau, physical AI technology lead at AWS’ Generative AI Innovation Center.
Tutor’s approach to advanced robot learning set it apart from other startups in the robotics and physical AI space, Simoneau said.
“What they were trying to do was build a brain for robots that could adapt to different environments,” he said. “And that generality was a pretty interesting thing.”
The main goal of the scholarship, which has expanded globally this year, is to advance cutting-edge technologies. However, some of the participants have matured and scaled to the point where they can sell their services to customers through the AWS partner network, Simoneau said. Rollout in Amazon warehouses isn’t the ultimate goal, he said, but could be a byproduct.
“Having a cloud partner that can support massive amounts of computation is critical,” Gruenstein said. All Tutor bots are connected to AWS.
Gruenstein said he thinks of Tutor as two halves: one focused on developing fundamental robotics research and figuring out how to build robot models and software, while the other focuses on deploying robots in factories and warehouses.
“I don’t think there’s any other company out there that’s building the robot brain and then deploying it in the real world and doing productive work in that manufacturing context or in that logistics context,” he said.
Tutor deploys its Cassie robot to US manufacturing and logistics facilities
While the Sonny robot is still in development, Tutor has deployed its Cassie robot for autonomous case picking and palletizing with US customers.
The 2,000-pound industrial robot can handle boxed products and materials, among other tasks, for a wide range of customers, Gruenstein said. He declined to provide details on the cost and number of Cassie deployments. early adopters from the food manufacturing and logistics sectors attended the tour and said they were already seeing the benefits of their pilot experiments.
Paul Baker, CFO of Productiv, a third-party logistics provider, said Tutor’s robot performs at or better than a human in its Dallas warehouse. The company is testing about 15 more robots that perform at subhuman levels, performing tasks of high variability and dexterity, with the goal of improving speed and efficiency.
“We are working with groups, including the Tutor, to try [improve dexterity] as fast as possible,” Baker said. “This is kind of the next frontier, and that’s why they have the data center running.”
Tutor Intelligence’s Cassie robot autonomously picks up and stacks boxes on a pallet as part of a demonstration.
Nathan Owens/Manufacturing Dive
Jeff Pulley, director of facilities for BetterBody Foods, maker of peanut butter powder, avocado oil, chia seeds and other health-focused products, said the company in July added two Tutor palletizing robots at its Lindon, Utah, facility and one at a Greenfield, Mass., factory.
Since implementing the robots, Pulley said BetterBody Foods has saved 36 percent compared to the cost of a worker who would normally have to load and unload boxes from pallets. He said the company hopes to add three more Cassie robots in their facilities for the next year and a half and eventually add Sonny robots when they are ready.
“I think robotics and automation is … where everything is going. If you want to delay it, you’re slowing down your own progression,” Pulley said. “So then [you] might as well accept him and learn from him.”
