Episode Summary
Executive Summary: Sunday co-founders Tony Zhao and Chang Chi argue home robotics is entering the “pre-ChatGPT” to “ChatGPT” transition: scalable methods now exist, but productization still requires massive data, robust hardware, and full-stack integration. They describe how imitation learning, diffusion policies, and glove-based data collection enable generalization beyond lab demos, and outline a near-term plan to beta test Memo in homes in 2026.
Main Topics: State of AI robotics and why this moment matters (Priority: 5/5): The guests frame robotics as being between a technical breakthrough and a mainstream product moment: the core recipe for manipulation seems known, but scaling data and systems is the remaining challenge before consumer-grade robots become viable. Diffusion policy, ACT, and imitation learning breakthroughs (Priority: 5/5): They explain how diffusion policies stabilized imitation learning by handling multimodal behaviors, and how transformers plus action chunking improved dexterous manipulation once enough high-quality data existed. Scaling data collection beyond teleoperation (Priority: 5/5): A major theme is moving from lab-bound teleop to scalable in-the-wild collection using GoPros and the UMI glove, which enabled large datasets, broader participation, and better generalization. Building Sunday as a full-stack robotics company (Priority: 4/5): The founders emphasize that robotics requires integrated work across hardware, controls, data pipelines, training, operations, and supply chain, making an in-house full-stack approach essential. Product philosophy for a home robot (Priority: 4/5): They discuss designing a cheap, safe, cute, and capable robot for everyday homes, with simplifications like a three-finger hand and compliant low-cost actuators to balance usability and manufacturability. Demos, generalization, and what robotics launch videos actually prove (Priority: 4/5): They caution against over-reading demos, stressing that one successful task does not imply broad capability. Their demos are presented as evidence of long-horizon manipulation and zero-shot generalization in real homes. Timeline, cost, and future of chores (Priority: 4/5): Sunday plans a 2026 beta program and aims to bring costs under $10K at scale, with the broader vision of making household labor marginally free and reducing mundane chores.
Key Arguments: Robotics is at a scaling inflection point: the field has the right methodological direction, but it needs much more data and better systems to achieve product-level performance. Classical robotics was slow because it relied on hand-engineered modules tailored to each task and environment, causing progress to reset with every new problem. Diffusion policies improved imitation learning by allowing multiple valid behaviors for the same observation while preserving training stability. Transformers became effective in robotics only after data collection quality improved enough to support dexterous, high-diversity demonstrations. Action chunking matters because robots should predict trajectories over longer horizons rather than react every millisecond like a low-level controller. Teleoperation is too lab-bound and unintuitive for scaling; glove-based data collection makes it possible to gather large, natural, in-the-wild datasets. The UMI glove approach can convert human hand motion into useful robot training data from only video and motion tracking, reducing dependence on expensive teleop rigs. Data quality becomes more important, not less, as scale increases; automatic calibration and failure detection are required to make large robotics datasets usable. Simulation and RL are promising in some robotics domains, especially locomotion, but manipulation is harder to simulate than to demonstrate, so imitation learning currently scales faster. Home robots must be designed for safety, low cost, maintainability, and broad everyday utility, not just industrial precision. A full-stack company is necessary because hardware, data collection, and model training continually influence each other, requiring rapid in-house iteration. A successful consumer home robot must be general enough to work in unseen homes without requiring users to collect new data after unboxing.
Data Points: Target number of robots in homes: more than 1 billion within a decade - The opening vision for the future of home robotics Company headcount: 30 to 40 people - Sunday team size at the time of the interview Earlier team size: 8 people - Team size toward the end of 2024 before scaling up Data collected for espresso task: 1,500 video clips - UMI-based data collection for the espresso cup serving task Training data scale: almost 10 million trajectories - Current in-the-wild data collected across tasks Glove versioning: V0 to V5 with around 20 iterations each - Hardware/data-collection device evolution over time Glove users: more than 500 people - Number of people using the gloves in the wild Prototype robot cost: $6,000 to $20,000 - Current in-house prototype cost range Expected scaled material cost: likely under $10K - Projected cost after scaling to a few thousand units Beta program timing: 2026 - Planned home beta rollout before mass consumer launch Academic research spend: about $200,000 - Reference to total research spend translating into notable data-scale impact Paper timing gap: one month or two months - Relationship between ALOHA and ACT paper release timing
Pivotal Quotes: "Nobody wants to do their dishes. Nobody wants to do their laundry." — Tony Zhao: Opening motivation for why home robots will be broadly desirable "we feel like we are between the GPT moment and the chat GPT moment" — Tony Zhao: Explanation of robotics being at the stage where a scalable recipe exists but the consumer product is not yet fully realized "we want the robot to feel like it's out of a cartoon movie" — Tony Zhao: Design philosophy for a friendly, non-threatening home robot
Implications: Home robotics may be closer than expected, but the winners will likely be companies that solve data, hardware reliability, and end-to-end product design together. For consumers, the near-term promise is relief from chores; for the industry, the bar is no longer demos, but robust generalization in real homes.