500 Gloves, 10 Million Trajectories, and a $10,000 Home Robot

Memo, Sunday Robotics' home robot, opens with a line the founders clearly mean: nobody wants to do the dishes or the laundry, and people would rather spend that time with family. If the machine is cheap, safe, and capable, Tony Zhao and Cheng Chi say everyone will want one, and they picture more than a billion of them in homes within a decade. Sarah Guo of *No Priors* sits down with the two co-founders to ask how much of that future is still research, how much is product, and how soon anyone can actually buy it.

Sunday calls Memo the first general-intelligence personal robot, aimed at chores rather than office productivity. Zhao is first author on Stanford's ALOHA and ACT papers; Chi is first author on Columbia's Diffusion Policy and UMI. Those four pieces of work, produced on what Guo estimates as about $200,000 of academic funding, largely set the last three years of robot manipulation. The company now has to turn that stack into something that ships.

Keep three numbers in mind for the rest of the conversation. More than 500 people in the wild now wear Sunday's data-collection gloves. The team has gathered almost 10 million trajectories — not two-second "pick up a cup" clips, but long sequences with walking, navigation, and full tasks. And once production reaches a few thousand units, they expect material cost, and a selling price nearby, to land under $10,000.

Guo wants a reality check. Classical robotics has not been a field of great optimism, and yet people now talk about a foundation model for robots, or a ChatGPT moment. Chi's answer is precise: the field sits between the GPT moment and the ChatGPT moment. In language-model terms, GPT is a technology and ChatGPT is a product. The recipe looks like it can scale, but it has not been scaled far enough to yield a great consumer product.

Academia, he says, has reached consensus on the method for manipulation, and everyone is talking about scaling up. There are signs of life in the algorithms people have picked, but nobody knows what happens when the data jump looks like GPT-2 to GPT-3. The trend is still clear. There is no reason to believe robotics will refuse the pattern of every other AI field: more data, better performance.

Zhao explains why the old process was so slow. Classical robotics used a sense-plan-act stack with human-designed interfaces between modules, rewritten for every task and every environment. In academia that meant one paper per task: design the task, design the environment, design the interfaces, do the engineering, then throw the code away and start over. Industry did the same thing with bespoke hardware and software per application, and there was no synergy between systems. It felt like running in loops, building one machine and then the next, which is why decades of work moved so little.

Sunday's technical base is three papers in sequence, starting with Chi's Diffusion Policy. Imitation learning is the most intuitive way to use machine learning for robotics: collect paired observation and action, train with supervised learning, and have the robot do the same thing. Researchers told Chi the paradigm was finicky — the specific person had to collect the data, with exactly one way to do each motion, or training would diverge and the robot would behave strangely. Diffusion models let a policy capture multiple modes of behavior for the same observation without losing training stability. That unlocked collection from several people, even untrained people, with results that still held.

ALOHA and ACT landed a month or two apart, which is how Zhao and Chi met: reading each other's papers and talking on Twitter while Chi was still at Columbia. Typical teleoperation used a VR headset and felt disconnected, a person waving in free air while the robot moved with lag. ALOHA cut that delay until it felt more like a video game, which is how you actually capture dextrous data. ACT then threw transformers at those datasets. Robotics had been stuck on three-layer MLPs and convnets that got worse as they got deeper; with strong data, a transformer worked. Action chunking — predicting a trajectory instead of a twitch every millisecond — mattered for the same reason humans do not refocus every millisecond. Motion became consistent, and overall performance jumped.

UMI is the company's seed. After ALOHA and Diffusion Policy, both founders were still doing teleoperation, and Chi found it limiting: a PhD student needed a couple of hours to set up a rig, which locked collection in the lab. A product has to work in unseen homes, so the data has to come from the wild. He asked whether you could collect robot data without a robot. Once imitation learning is just paired video plus hand motion and finger opening, a GoPro can record all of it. He 3D-printed a handheld UMI gripper, and two weeks before the paper deadline three PhD students took it everywhere, including restaurants while waiting for a table. They gathered 1,500 clips of serving an espresso cup, then one of the larger datasets in the field, and trained the first end-to-end model that generalized to unseen environments. They could push an arm around the Stanford campus and have it serve a drink anywhere.

Zhao was there, and that campus demo is when he thought they should start a company. The exception was direct sunlight. Those two weeks had been rain, so the training set had no sun, and the policy failed outdoors. The lesson became a company habit: distribution matching. For a robot to work on a sunny day, the training data must include sunny days. Guo notes they spent on the order of $200,000 across the academic work and still led on data scale translated into capability. Robotics has no internet-text dump to train on, the field is blocked on data, and the live debate is whether to scale via world models, simulation, or teleoperation. Sunday wants a path that is both high quality and scalable.

The company started as the two of them in Chi's apartment, with a robot clamped to his desk. By the end of 2024 they were eight people; they are now around 30 to 40. Starting a company, Zhao says, let them hire people they love working with across mechanical engineering, supply chain, software, and controls, and build a product rather than a demo. The mission is a home robot in everyone's home. Plenty of AI tries to make you more efficient at work; not enough of it takes the mundane, uncreative chores that have nothing to do with being human.

If capable robots are going to be everywhere, people will see them every day, so the surface-level design question is what that should feel like. The team wants Memo to have a face — a cute, friendly one — so the machine feels like it walked out of a cartoon rather than like a Terminator doing the dishes. The obvious alternative is a full humanoid, and Zhao's answer is speed: whenever simplification gets a useful robot out sooner, they simplify. Memo's hand combines three fingers into one, because people already use those fingers together to grasp a handle or open a dishwasher. Paying 3x for separate fingers did not buy enough benefit. The constraint underneath every cut is that the robot should eventually do all the chores, while staying low-cost and easy to repair.

The actuator philosophy shifted for the same reason. Industrial robots are fast, stiff, and precise because they are blind: they follow a programmed trajectory and cannot react to what they see, so mechanics have to carry all the accuracy. AI gave robots eyes, Chi says, so the machine can correct its own mechanical error on the fly. That opens a different design space: cheap, compliant, imprecise actuators, with the model supplying enough accuracy for home tasks and the hardware staying inherently safer around children and pets. A fast, rigid machine was never going to be welcome in a house.

Asked where they sit on the GPT-to-ChatGPT timeline, Zhao says 2026 is a beta year: robots, of several prototype kinds, go into people's homes so the company can learn whether owners want to talk to them, have them teach kids, or simply take chores. Internally the bar for a minimum consumer product is extreme safety, extreme capability, and low cost. What they know now, and did not at the start, is that adding data improves both dexterity and generalization, and the company is the cross product of those two axes. A generalization demo — picking real metal forks off ceramic plates at high success rates — worked more easily from scale than they expected. Chi's version is less glamorous: scaling a robot system was long and painful, with mechanical failures, reliability gaps, and data-quality problems they had underestimated. Compared with teleoperation, the glove system was much harder to scale; once it scaled, it was powerful and repeatable. Guo's summary lands: it was harder than they thought, and they are further than they thought.

Early on they had a nervous joke: if they built the glove, someone could take it and copy the whole thing. The path taught them otherwise. The data-collection system, the robot and control stack, the filtering and cleaning pipelines, and the training pipeline all have to iterate together, several loops in, which is hard to imagine without a full-stack team under one roof. The glove in use now is V5; each generation from V0 had around 20 internal iterations, or about a hundred in total. More than 500 people use those gloves in the wild, which means everything that can go wrong does. If assembly instructions are vague, people assemble the device in creative ways, and creativity is the enemy of a sensor that has to be extremely precise.

Relative to the rest of the field, they are at almost 10 million in-the-wild trajectories, and those are long-horizon: walking, navigation, then the task, not an isolated cup grasp. They had worried glove data would be high quantity and low quality next to teleoperation, because teleop uses the same hardware and software at train and test time — a perfect distribution match. The opposite happened. The glove form factor encourages more dextrous, more natural movement, which shows up as more intelligent behavior in the model. After a full engineering pass they no longer see a quality gap. The remaining hard problem is converting human data one-to-one into robot data: the camera frame contains a person, and the model must not be able to tell.

On reinforcement learning versus imitation, Zhao is blunt. RL is powerful and much less sample-efficient; it shines where simulation is easy, which is why it works for locomotion — rigid-body contact between feet and a somewhat flat ground. The behaviors you actually want, like retracting a foot after a bump, are hard to learn from demonstrations, but the world is cheap to simulate. Manipulation flips the difficulty. Grasping a transparent cup of orange juice is ridiculously expensive to simulate, from finger deformation to juice ripples to the rendering the policy will see, while the behavior itself is simple: get the hand in front of the cup and close with the right force. If you had a perfect world simulator you could cure cancer in it; the real question is which parts of robotics are cheap to simulate. That is why imitation has moved faster on manipulation. Scaling also taught Chi that data quality really matters in a way he already believed but had not felt, so the company calibrates every glove before it ships and runs software that flags a broken unit without a human staring at the logs. The two technical problems still in front of them are the training recipe at this new scale, and hardware that the learning team keeps breaking. Research itself only started about three months ago, after they built infrastructure, because data-scarce eras are too good at producing cute ideas that do not scale.

Guo asks the uncomfortable commercial question: when can people buy a home robot? Zhao is excited because the office is already full of prototypes they want out in the world. The next step is a 2026 beta: selected people who sign up get a real robot that starts doing chores, which is how Sunday will watch humans interact with the machine and learn which jobs people actually want done. Mass shipping waits on an "incredibly high standard" for consumer experience — highly reliable, capable, and cheap. Whether that is 2027 or 2028 depends on the beta. It is definitely not a decade away.

In-house prototypes currently cost somewhere between $6,000 and $20,000, and the spread is not a better actuator. Every prototype uses the same low-cost actuators. The variable is cladding: at low volume the shells run a few thousand dollars because they are CNC-machined and hand-painted. At a few thousand units those parts become injection-molded and, Zhao says, dirt cheap. Material cost likely falls under $10,000 at that scale, and the selling price would sit somewhere around it.

Five years out, if home robots are ubiquitous, Chi's picture of daily life is almost embarrassingly specific. His sink usually holds four or five somewhat dirty dishes that smell a little, and after a long day that is a miserable thing to walk into. The world they want is cleaner. Guo's version is economic: the marginal cost of labor in homes goes to zero. First the house gets clean.

Guo points at the genre everyone has watched: launch videos that look astonishing, including Tesla Optimus serving drinks at a bar years ago, with no product behind them. Zhao's method is four words, and a nod to the show: make zero assumptions. First, is the motion autonomous or teleoperated? Read what the poster actually said. Second, if a robot hands one cup to one person, did they also show a slightly different color cup going to someone else? If they did not, assume it can only pick up that single cup for that same person. Viewers project human instinct onto the clip — if it can hand a cup, it can do the dishes — and that imagination is part of why robotics is fun, but it is a bad way to score a demo. Index only on what is shown; that is likely the full scope of the task. He also counts interactions, because each contact is a chance of failure, and a longer sequence is simply harder.

Sunday's own demos come in three categories. The first is a messy table: dump leftovers into food waste, load dishes into the dishwasher, then operate the dishwasher. What makes it hard is the mix of fine manipulation with long-horizon mobile work, going up and down through a full routine. The second is force. One hand picks up two wine glasses — transparent objects — and has to seat them precisely in the dishwasher. Squeeze a little harder and a glass breaks; push in the wrong direction while loading and it shatters. They broke a lot of glasses getting it to work. Chi's point is that many home tasks are high-stakes in a particular way: the job is not recovering from mistakes, it is not making them in the first place.

The third category is generalization, and the closest thing they have to a real product. They booked six Airbnbs, rolled the robot in zero-shot, and tried two jobs: circling a table to collect utensils into a caddy, and grabbing a plate into the dishwasher. They collect no data on entry, which is the point — a customer should not have to record a dataset to unbox the machine. The silverware is whatever the home already owns, grasps need a few millimeters of precision, reflective metal confuses perception, and some tables are glass. It still works, they say, because more than 500 collectors have already seen so many glass tables. The dexterity peak is an espresso machine and folding socks, both of which need fine force that teleoperation cannot capture. Current teleop has no good force feedback, so "your hand is numb": you may be applying huge force on the robot without knowing it, and the data becomes a pile of aggressive motions. A glove is different. The person folding a sock can feel the tension, two fingers close into a force closure, and the force is whatever it actually was. They believe they are the first in the industry to fold socks and run an espresso machine with an end-to-end model.

To scale the company they are hiring full-stack roboticists, and people who want to become them. Robotics is multidisciplinary: a little mechanical, a little electrical, a little code, a little data, or you cannot optimize the system. They have already turned software engineers and mechanical engineers into roboticists. Chi is the walking example. He did not write code before college, starting in mechanical and electrical design, then learned programming when he realized the bottleneck was how the robot moved, then found machine learning near graduation. Whatever you can imagine about robotics, consumer products, and machine learning, he says, you can find at Sunday, because the company is the whole experience, not just the software or the hardware.