Summary: Hello Robot began selling Stretch 4 in May at $29,950, and its opening run of 200-300 units sold out. The home-assistance robot deliberately ships limited autonomy with human control, while a quadriplegic user reported tasks such as making a protein shake falling from nearly two hours to minutes. Research use also supports the company’s data thesis: zero-shot Robot Utility Models completed five mobile manipulation tasks with 90% success in unseen environments.
Martinez, California sits about 45 miles outside Silicon Valley, and that is where TechCrunch reporter Tim Fernholz headed on June 4, 2026, to find Hello Robot assembling the Stretch 4, the fourth generation of its home assistance robot. Aaron Edsinger, formerly Google’s robotics director, and Georgia Tech professor Charlie Kemp started the company in 2017; this May the new machine went on sale — $29,950 on the official site, tagged “available now” — and the opening run of 200-300 units has already sold out.
Home robotics is having a strange year. 1X has collected 10,000 reservations for its Neo humanoid without delivering a single one, and The Bot Company is being sued by a San Francisco Airbnb owner after its robot allegedly damaged furniture, appliances, and tiles. Hello Robot runs the other direction: skip the full-autonomy promises, ship hardware first, and let real operating hours start piling up.
Meet the Stretch 4
Nothing about the Stretch 4 pretends to be a humanoid. There is a torso loosely shaped like a human upper body, a head bristling with sensors, one telescoping arm ending in pinchers, and an omnidirectional base on wheels. Engineer Blaine Matulevich jokes that the battery-status lights ringing its eyes make the robot “look angry.” What actually sets it apart is how it gets to buyers: the whole robot fits inside a cardboard box and travels by UPS or DHL — no crates, no installation crew at your door — which amounts to a real concession on hardware cost.
Chinese rivals price theirs lower, but Edsinger’s reply is that the sticker is not the total: robots sold cheap leave out the sensors and software that make them genuinely useful, and the buyer pays for those gaps somewhere else. On its site, the company points its product line at three audiences — home assistance for people with severe mobility impairments (an in-home pilot program is already open), an open research platform “for researchers, by researchers,” and feasibility pilots inside enterprises.
Humans in the Loop by Design
Out of the box, the Stretch 4 can only do so much on its own, and that is on purpose. “Being in control is a feature — it’s desired to be embodied in the robot,” Matulevich says. To explain the safety-first posture, Edsinger reaches for Waymo’s playbook on self-driving: make predictability the product instead of staging the sci-fi demo. Investors are drifting the same way. As a Bullhound Capital report puts it, in robotics “the moat isn’t just IP, but accumulated operating hours under real-world liability.”
How One Quadriplegic User Lives With It
Nothing makes the value case better than how Keith Platt gets through a day. Platt, a company board member and investor, has been quadriplegic since 2021. From a voice-operated iPhone app he drives Stretch — the robot covers part of each task with its own navigation, and he takes direct control for the rest. Mixing a protein shake once cost him close to two hours; it now takes minutes. The robot also fetches his reading glasses and helps with toothbrushing. Wider adoption, he says, would be “life-changing” for families like his.
The Real Moat: Data
Hello Robot stays out of the foundation-model business. Its role is to be the platform other people attach AI brains to, and researchers have put that role to heavy use. Mahi Shafiullah finished his NYU PhD on a Stretch 3, then helped NYU, Meta, and Hello Robot build Robot Utility Models: zero-shot robot policies that completed five mobile manipulation tasks — opening cabinet doors and drawers, picking up napkins and paper bags, reorienting fallen objects — with a 90 percent success rate in environments they had never seen. The policies run on-device, an external multimodal LLM verifier catches failures and triggers retries, and the entire stack shipped open: MIT-licensed code, a dataset of roughly 5,509 trajectories across 180 environments, and The Stick V2, a 3D-printable data-collection tool. His verdict on home robotics: “The algorithms may be there, but the data is not, and data is actually like 80% of the ingredient that matters.” Hardware, in his words, stays “actually abysmal” for home use — heavy limbs, high energy demands, damage-prone mistakes.
Product teams can squeeze the lesson into one line: the home robot of 2026 is not a matter of waiting for one do-everything model. It is the discipline of shipping limited autonomy plus dependable human control, at a price and in a form factor users can actually reach — because the robots already in homes are the ones collecting the data the next generation will train on.