Martinez, California, is about as far as you can get from Silicon Valley and still be in the San Francisco Bay Area. Perched on the northeast edge of the bay, the small city is home to Hello Robot, a startup that itself is about as far as one can get from the maximalist promises of its robotics rivals 45 miles south.
Hello Robot released the fourth iteration of its home assistance robot, Stretch, last month. And you might stretch to call it a humanoid robot. While Stretch boasts a vaguely human torso and sensor-studded head, its telescoping arm has a pair of pinchers, and it rides around on a heavy, omnidirectional wheeled base.
When Stretch’s batteries run down, lights around its “eyes” glow — “it looks angry,” Blaine Matulevich, an engineer at the company, jokes.
Hello Robot, founded in 2017 by CEO Aaron Edsinger, a former director of robotics at Google, and CTO Charlie Kemp, a professor at the Georgia Institute of Technology, is not building a foundation model or promising to take over every job a human can do. Hello Robot developed Stretch to do something many other robots aren’t doing: Working in real homes, with real people, at a time when most are behind glass in laboratories.
This is vital. While the latest advances in artificial intelligence promise more capabilities for robots, there is a dearth of useful training data. And while simulation is improving, investors are increasingly focused on deployment.
“Companies that deploy first accumulate site-specific recovery loops and workflow tolerances that no competitor can buy or synthesize,” Bullhound Capital wrote in a report on the sector published last week. “In robotics, the moat isn’t just IP, but accumulated operating hours under real-world liability.”
A different kind of embodiment
Keith Platt, an investor in Georgia who now sits on Hello Robot’s board, invested in the company after taking on Stretch as a housemate. Platt became quadriplegic in 2021, only able to control parts of his shoulders, his neck and his head. He began exploring adaptive technology, and in 2024 started working with Hello Robot, which has an occupational therapist on the team to support its work with Platt and other people with similar conditions.
Platt controls his Stretch using a voice-operated iPhone app; he can task it to autonomously move to somewhere in his house, then take over direct control to manipulate objects and perform tasks. One deceptively simple project has been figuring out how to get Stretch to serve him a protein shake for breakfast, which normally requires the assistance of another person.
“When we first started out with that activity, it took me independently — no one there — took almost two hours,” Platt told TechCrunch. “But I was gonna stick with it. It got down to where, within a few minutes, I could drink the whole shake and put it back on the counter.”
Being dependent on people is a real challenge, both physically and emotionally, Platt says. Anything he can do to regain independence — like putting on or taking off his reading glasses, or brushing his teeth himself — “is huge.” Not just for him, but for the people who care about him.
He predicts it would be “life-changing” for families if robotic assistants could enable people with mobility challenges to be able to safely spend a day at home, allowing their family members to work independently or leave the house without hiring a professional caregiver.
Stretch comes from the factory with limited autonomy; focusing on having a human in the loop is intentional. “Being in control is a feature — it’s desired to be embodied in the robot,” Matulevitch said.
And, Platt points out, he doesn’t worry about Stretch falling over if it suffers an error.
Hardware is hard
For all the money flowing into startups designing brains for robots, their bodies still leave a lot to be desired. While components are getting cheaper, the state of the art still delivers heavy limbs that require high-energy, active balancing. A robotic hand and arm weighs much more than a human’s, and physics is unforgiving.
When robots make mistakes, they damage things around them. One startup, the Bot Company, is being sued by a San Francisco Airbnb owner who says the company rented his apartment to work on its robot, which scratched furniture, broke appliances, and chipped bathroom tiles.
“The state of hardware today is actually abysmal from the perspective of, ‘I want to have robots in my parents’ place,’” Mahi Shafiullah, a postdoc working on robotic hands at the University of California, Berkeley, told TechCrunch. He recalled industrial robots in his lab accidentally punching through a plastic kitchen play set they were supposed to carefully manipulate.
Shafiullah ultimately came to use the third generation of Hello Robot’s Stretch as part of his PhD research at New York University. Models he helped develop with Stretch won the best demonstration prize at least year’s Computer Vision And Pattern Recognition (CVPR) conference.
Hello Robot doesn’t promise that Stretch will have the complexity or capability of the humanoid robots that enamor the Valley, but its simpler design could make it more powerful. Edsinger compares his company to Waymo, which became the leading purveyor of self-driving cars by focusing on safety first (although the money helped).
One leader in this field, 1X, was the subject of significant attention last year when it unveiled a humanoid robot, Neo, that people could buy to perform chores in their homes. The company says that it sold out of the 10,000 Neos it plans to build this year, but as of yet, none have actually been delivered.
“Hello Robot has been really cautious and really caring about this problem, because I think they’re designing it to be around people first,” Shafiullah said. “And then they’re thinking about, where are the capabilities that they can fit in within those limitations?”
Homeward bound
Stretch 4 costs an affordable-for-a-robot $30,000, which is a bit more than robots from Chinese manufacturers, although Edsinger notes that those often don’t come with sensors or software included, add-ons that ultimately drive up the price. He expects to manufacture between 200 and 300 at the company’s Martinez headquarters, with the first run already sold out.
Edsinger wants to keep the robot accessible to hackers and researchers on low budgets. One design criteria for Stretch is that it has to be shippable in a cardboard box via UPS or DHL— once wooden crates and installation teams are required, costs go up and accessibility declines.
Hello Robot’s customers include researchers who use Stretch to test out increasingly sophisticated AI brains, enterprise customers who are testing Stretch’s utility in settings like data centers, and people working to develop in-home aides for people with disabilities.
The combination of the robot’s comprehensive sensor suite, physical capabilities, and safe operations could make it a candidate to fill out the hopes of physical AI believers.
“The algorithms may be there, but the data is not, and data is actually like 80% of the ingredient that matters,” Shafiullah said.
Having a robot that can safely collect that data is another step forward. And Hello Robot intends to keep iterating. The lessons from the roll-out of Stretch 4 promise to feed into the company’s next bot, which could drive down the price and increase the capabilities enough to realize a vision of robot-human collaboration at home.