For decades, we’ve imagined robots in our homes—helping with chores, tidying up, and quietly taking care of the everyday. The vision predates The Jetsons, and yet households still rely on manual labor. The reason is straightforward: the home is one of the hardest places to automate. It is unstructured, constantly changing, and full of edge cases. Tasks require dexterity, judgment, and contextual awareness. Systems must operate safely around people, at consumer-grade cost.

To make home robotics work, three things must come together: capable hardware, intelligence that can handle the long tail of real-world situations, and a scalable way to capture the distribution of human behavior. That last piece has been the biggest bottleneck. It’s hard to capture the messy data that reflects how people actually live. As a result, robotics has often remained stuck in prototypes and flashy demos.

Sunday Robotics starts from a clear thesis: generalizable home robots require large-scale, high-fidelity demonstrations drawn from real households. Historically, collecting this kind of data has been expensive and operationally complex. Lab environments and simulation have been useful, but they fail to capture the long tail of domestic life: unusual layouts, clutter, pets, and unpredictable object placement. Teleoperation rigs and robot twins often cost tens of thousands of dollars and require trained operators in controlled environments. Meanwhile, egocentric video and other lightweight data sources do not translate cleanly to robot control because the human and robot embodiments differ. The result has been datasets that are small, narrow, and biased toward artificial lab conditions—far from the messy distribution of real homes where these systems ultimately need to work.

Sunday’s technical breakthrough is a scalable demonstration pipeline built around a low-cost (~$400) glove system that mirrors Memo’s kinematics and control stack. A distributed network of U.S. based “Memory Developers” uses these gloves to record themselves performing everyday tasks in their own homes. Importantly, these are all U.S. households. It’s common for folks to enter lower cost countries for labor, but those environments won’t resemble end-user homes that Memo is designed to operate in. Today, well over 500 Memory Developers contribute daily demonstrations, and the company is on track to 5x data operations this year.

Because the demonstrations are captured through hardware that reflects the robot’s embodiment, the data transfers cleanly into the robot’s learning pipeline. This alignment also reduces the amount of data required to train new skills. Each new Memory Developer adds another slice of the real-world distribution – different layouts, different household items, and different ways to approach the same task. As the network grows, the dataset naturally expands across the long tail of domestic variability, creating a data flywheel grounded in real environments. Over time, this coverage across homes and tasks establishes a practical path toward generalizable household competence.

Their first robot, Memo, is the product expression of this approach. It is designed to perform everyday household tasks reliably under practical constraints, which means prioritizing robustness over feature breadth to start. Memo is vertically integrated by design. By owning the full stack, Sunday can iterate quickly, diagnose failures end-to-end, and tune the system for reliability in real homes. As Memo is deployed, every interaction becomes training data, feeding back into the same learning pipeline that created it and steadily expanding what the robot can do.

Founders Tony Zhao and Cheng Chi first met on X after coming across each other’s papers, published about a month apart. Tony’s ALOHA introduced a system for imitation learning directly from real demonstrations collected through a custom low-cost teleoperation interface, along with a novel algorithm – Action Chunking with Transformers (ACT) – to make learning from those demonstrations easier. By predicting short sequences of actions (“chunks”) rather than individual timesteps, they could learn difficult manipulation behaviors from only minutes of demonstration data. Cheng’s Diffusion Policy showed that you can outperform prior approaches for generating robot action trajectories using diffusion, which tends to handle multimodal and high-dimensional action spaces better. He uses a modified form of the diffusion policy algorithm in his UMI paper, which introduced a low-cost gripper system capable of capturing high-fidelity manipulation data. Across these projects, a pattern emerges: Tony and Cheng have a history of engineering low-cost hardware and scalable pipelines for collecting real-world demonstrations, and their contributions have shaped how the field approaches learning from human data.

What distinguishes Tony and Cheng is not just their research output but how they approach building. They have a rare ability to translate ideas into working systems in environments where constraints like safety, cost, and reliability really matter. They combine deep technical rigor with strong product instincts. They are not pursuing robotics as an academic exercise. They are builders first, focused on creating the most delightful, safe, and useful home robot.

A home robot has to coexist with people. That means thinking about how it moves, how it communicates intent, and how it earns trust over time. Small design decisions matter. If you’re going to look at a robot every day, it probably should have eyes. Memo does—and it’s a genuinely cute robot.

We invested in Sunday Robotics because they are taking the most grounded path toward a category that has historically attracted grand visions and few real deployments. They are making deliberate tradeoffs to get useful robots into homes sooner, understanding that adoption is driven by reliability, trust, and everyday value. The home represents one of the largest untapped labor markets in the world, and the first company that learns fastest from real homes will define the category. Sunday Robotics is positioned to do exactly that. The team is exceptional, bringing experience from Tesla, DeepMind, Waymo, OpenAI, Meta, and Apple. They are actively hiring across software and hardware engineering, robotics, marketing, operations, and other key roles.

Home robotics won’t arrive through a single breakthrough. It will emerge through systems that are useful enough to deploy, learn from real-world usage, and steadily expand what’s possible.

We’re excited to partner with Sunday Robotics as they work to bring capable, reliable robots into everyday homes – so you can spend your Sundays doing something different.

For decades, we’ve imagined robots in our homes—helping with chores, tidying up, and quietly taking care of the everyday. The vision predates The Jetsons and yet households still rely on manual labor. The reason is straightforward: the home is one of the hardest places to automate.

Sunday Robotics is tackling this challenge with Memo, a vertically integrated home robot trained on large-scale, high-fidelity demonstrations captured in real U.S. households. By using a low-cost glove system that mirrors the robot’s embodiment, they collect rich, transferable data from a distributed network of “Memory Developers,” enabling Memo to learn reliable household skills across the long tail of real-world variability.

Founded by Tony Zhao and Cheng Chi—researchers behind ALOHA, ACT, Diffusion Policy, and UMI—Sunday Robotics combines cutting-edge imitation learning with pragmatic product thinking. Their focus is on reliability, safety, and everyday usefulness, building a data flywheel where every in-home interaction becomes new training data.

Sunday Robotics is hiring across software, hardware, robotics, marketing, and operations as they work to bring capable, trustworthy robots into everyday homes.

Key pillars of Sunday Robotics:
- Low-cost, scalable data collection via teleoperation gloves
- Learning from high-fidelity demonstrations in real U.S. homes
- Vertically integrated hardware and software stack for Memo
- Continuous improvement loop: every deployment generates new training data