Sunday Robotics: Scaling the Home Robot Revolution with Co-Founders Tony Zhao and Cheng Chi

Sunday Robotics Bets Home Robots Are a Scaling Problem

Sunday Robotics is building Memo, a general-purpose home robot, on the bet that manipulation has a workable recipe but lacks scale. Its founders helped create imitation-learning methods and data-collection tools that showed cheap hardware and ordinary people can gather robot data. Now they are running a large data pipeline and designing a safe low-cost robot. They aim for an in-home beta in 2026.

Takeaways

• Home robots are a scaling problem. Zhao and Chi say the manipulation recipe is broadly agreed, but nobody has scaled it into a great consumer product. They compare the moment to GPT-2, where more data may unlock unknown capability.
• Data pipeline came before research. Sunday Robotics built infrastructure first, now on version five of its data-collection glove with about 20 iterations per version. More than 500 people collect in the wild, producing almost 10m trajectories before research started three months ago.
• UMI collected robot data without robots. Chi's 3D-printed GoPro gripper captured paired observation and action. Three students gathered 1,500 espresso-serving clips in two weeks, one of robotics' biggest datasets, and trained the first end-to-end model that generalized to unseen environments.
• Imitation learning beat classical modular robotics. Classical sense-plan-act needed custom interfaces for every task, so work did not compound. Diffusion policy handled multiple behavior modes and stayed stable, while ALOHA and action chunking made transformers work and let untrained people collect data.
• Design targets safety, capability, low cost. The robot has a cute face and a three-finger hand. Its compliant actuators rely on AI eyes to correct errors. Prototypes cost $6,000 to $20,000, mostly cladding, but injection moulding at scale should push material cost under $10,000.
• Reinforcement learning wins locomotion, not manipulation. Chi says reinforcement learning will keep winning at locomotion but is far less sample-efficient than imitation learning and struggles where simulation is hard. Zhao says simulating manipulation is harder than learning it, so data quality matters more at scale.

Roadmap and hiring

• Sunday Robotics plans an in-home beta in 2026, with selected people getting a robot that starts doing chores.
• Commercial shipping is possible in 2027 or 2028, and Zhao says the home robot is definitely not a decade away.
• Zhao expects more than a billion home robots in people's homes within a decade.
• The company is hiring full-stack roboticists, including people who aspire to become full-stack roboticists.