Sunday Robotics: Scaling the Home Robot Revolution with Co-Founders Tony Zhao and Cheng Chi
Timestamps are as accurate as they can be but may be slightly off. We encourage you to listen to the full context.
Podcast Summary
This episode features Tony Zhao and Cheng Chi, co-founders of Sunday Robotics, discussing their journey from groundbreaking AI robotics research to building Memo, the first general-intelligence personal robot. The conversation explores how the field of robotics is experiencing its own "GPT moment," with transformative advances in diffusion policy, imitation learning, and data collection methods. (00:56)
- Main themes: The evolution from classical robotics to AI-powered general-purpose home robots, scaling data collection for robotic learning, and the technical challenges of building a commercial robot that can perform household chores autonomously
Speakers
Tony Zhao
Co-founder of Sunday Robotics, Tony is a leading researcher in AI robotics who contributed to breakthrough work in diffusion policy and ACT (Action Chunking with Transformers). His research helped establish the foundation for scalable imitation learning in robotics, moving the field from traditional teleoperation to more intuitive data collection methods.
Cheng Chi
Co-founder of Sunday Robotics and creator of the revolutionary UMI (Universal Manipulation Interface) system that enabled large-scale robotic data collection in the wild. Previously at Columbia University, Cheng pioneered methods for collecting robotic training data without requiring expensive lab setups, contributing to some of the largest manipulation datasets in robotics.
Key Takeaways
Scale Data Collection Outside the Lab
Traditional robotics required expensive teleoperation setups in controlled lab environments, severely limiting data collection capabilities. (07:06) The breakthrough came when Cheng realized that robotic data could be collected anywhere using a simple 3D-printed gripper and GoPro camera system. Within two weeks before a paper deadline, three PhD students collected 1,500 video clips by taking the gripper to restaurants and other real-world locations. This approach generated one of the biggest datasets in robotics and enabled the first end-to-end model that could generalize to completely unseen environments, like serving drinks anywhere on Stanford campus.
Embrace Multiple Modes of Behavior in Training Data
Classical imitation learning was extremely finicky because it required exact replication of demonstrations, with researchers having to personally collect all training data to ensure consistency. (03:31) Diffusion policy revolutionized this by allowing models to capture multiple ways of performing the same task while maintaining training stability. This meant multiple people, even untrained individuals, could contribute to data collection without causing the model to fail. The approach unlocks scalable training by embracing the natural variation in how humans perform tasks rather than forcing rigid consistency.
Optimize Hardware for AI, Not Industrial Precision
Traditional industrial robots are designed to be extremely fast, stiff, and precise because they blindly follow pre-programmed trajectories without perception. (14:24) However, AI-powered robots with vision can correct their own mechanical inaccuracies in real-time. This paradigm shift allows the use of low-cost, compliant actuators that are inherently safe but less precise. The AI algorithms compensate for hardware imperfections, enabling robots to achieve the necessary accuracy for home tasks while being mechanically safe and affordable for consumer use.
Build Full-Stack Integration for Robotics Success
Unlike software companies, robotics requires unprecedented integration across mechanical engineering, electrical systems, data collection, AI training, and operations. (27:42) Sunday Robotics learned that working with external partners was challenging because their standards of "good" were constantly evolving as they discovered new requirements. Building everything in-house allows for rapid iteration cycles where hardware failures immediately inform mechanical design improvements, and AI team requirements directly influence data collection device development. This full-stack approach is harder to build but essential for creating a cohesive robotic system.
Prioritize Data Quality Over Quantity at Scale
While scaling up robotic data collection, the founders discovered that data quality becomes exponentially more important at scale. (24:08) With over 500 people using data collection gloves in the wild, every possible failure mode will occur - from creative assembly mistakes to hardware failures. This required building extensive automated monitoring systems to detect problems without human oversight, developing precise calibration processes for each device, and creating repeatable quality control procedures. The lesson is that scaling data isn't just about collecting more - it's about engineering robust systems that maintain quality across diverse real-world conditions.
Statistics & Facts
- Sunday Robotics has collected almost 10 million trajectories of robotic training data in real-world environments, making it one of the largest manipulation datasets in the field. (19:02) These aren't simple pick-and-place tasks but complex, long-horizon sequences involving walking, navigation, and multi-step household chores.
- The company now has over 500 people using their data collection gloves in the wild to gather training data for their home robots. (18:41) This distributed approach allows them to capture the diversity of real-world environments and tasks that robots will encounter in homes.
- Current prototype robots cost between $6,000 to $20,000 to manufacture, with the expectation that scaling to thousands of units could reduce costs to under $10,000. (29:15) Much of the current cost comes from low-volume manufacturing processes like CNC machining and hand painting, which will shift to injection molding at scale.
