No Priors Ep. 141 | With Sunday Robotics Co-Founders Tony Zhao and Cheng Chi

In this talk

1. 0:56

   Robotics may be entering a scaling phase

   Tony compares robotics to the period between GPT and GPT-2: researchers have promising methods, but have not yet scaled them enough for a strong consumer product. He says older robotics work often relied on task-specific sense-plan-act modules that had to be rebuilt for each environment, limiting reuse and slowing progress.

2. 3:14

   Better demonstrations made learning more scalable

   Tony explains that Diffusion Policy can learn multiple ways to act from paired observations and actions, making training more stable and allowing more people to collect demonstrations. Cheng says ALOHA made teleoperation feel more like a video game, while action chunking helped robots produce consistent movements. With UMI, people used a GoPro and a handheld gripper to collect 1,500 espresso-serving clips in about two weeks, including data outside a lab.

3. 10:38

   Sunday is building the full system for home use

   Cheng says Sunday grew from two founders working in Tony's apartment to a team of around 30 to 40 people, bringing mechanical, software, supply-chain, and controls expertise together. The company aims to reduce household chores and simplify the robot's design where possible, including using a three-fingered hand. Tony says perception lets the robot use lower-cost, compliant actuators and correct for mechanical imprecision.

4. 16:43

   Data quality and integration are central challenges

   Tony describes scaling as a difficult cycle involving hardware reliability, data collection, filtering, cleaning, and model training. Sunday has iterated through six glove versions, with about 20 revisions per version, and Cheng says more than 500 people use the gloves in the wild. They report nearly 10 million collected trajectories and say automatic calibration and fault detection help keep that data usable.

5. 24:55

   Home trials will test reliability and generalization

   Cheng identifies training at scale and reliable hardware as major technical challenges, while Tony says Sunday built its data, operations, and training systems in-house. The founders plan a 2026 beta program, with a wider release potentially in 2027 or 2028; current prototypes cost $6,000 to $20,000, and they estimate scaled production could bring cost below $10,000. They describe demos involving dishwasher loading, transparent glassware, and sock folding, and advise judging robotics demos by what they actually show and how many interactions succeed.