Sunday has rebuilt Tesla’s data playbook for robotics

Signals Inbox·July 18, 2026·Robotics

Sunday says it now runs the largest data collection operation in robotics, built around lessons Perry Jia learned during six years at Tesla and then spent two years adapting for physical AI. The bigger signal is not another polished robot demo. Sunday is trying to industrialize the slow, expensive process that turns human chores, robot failures, and strict evaluations into reliable autonomy.

Q1What actually happened?

Sunday cofounder Tony Zhao said the company now runs the largest data collection operation of any robotics company. He said Perry Jia spent six years learning data operations at Tesla, then another two years rebuilding that playbook for robots. Sunday has also brought its evaluation system in-house, so researchers can focus on improving models instead of managing data pipelines.

Q2Why copy Tesla’s playbook?

Tesla learned that autonomy improves through a loop: deploy machines, find failures, collect the right examples, retrain the model, and test again. By January 2025, Tesla customers had driven three billion miles using supervised FSD. A young robot company cannot match that scale, but Sunday is copying the operating idea: treat data collection as a core product, not a side job for researchers.

Q3Why is robot data so difficult?

A language model can learn from billions of pages already online. A robot needs examples that connect video, touch, movement, force, timing, and the result of each action. Someone or something must physically create those examples. Every new object, room, grip, and mistake adds another case, which makes real robot data much slower and more expensive to produce.

Q4How is Sunday collecting it?

Sunday uses sensor-equipped gloves that cost roughly $200 to build and mirror the shape of Memo’s hands. People perform everyday tasks while the gloves record movements that can be translated into robot actions. Sunday then uses its internal robot fleet to attempt those skills, study failures, and improve the model. It is building both the demonstration factory and the correction loop.

Q5How large is large in robotics?

The public benchmarks show how young this field still is. Open X-Embodiment combined data from 22 robots and covered more than 160,000 tasks. AgiBot World later published more than one million trajectories across 217 tasks. Another augmented dataset reached 4.4 million trajectories. Those sound large, but they remain tiny next to internet-scale text or video datasets, especially once you divide them across different robots and environments.

Q6What does Sunday get from owning evaluation?

It can measure whether a robot skill is truly reliable instead of selecting one successful attempt for a video. Sunday recently proposed a standard called a Solve, which states the task, environment, objects, extra training, and human help involved. That matters because a robot folding one familiar shirt once is very different from folding unfamiliar laundry in a home with more than 99% success.

Q7So what is the real competitive tension?

The robotics race is shifting from who can make the best demo to who can improve fastest after thousands of failures. Figure, 1X, Tesla, AgiBot, and research groups are all chasing more real-world experience. Sunday is betting that its moat will be the machinery behind learning: cheap human demonstrations, an internal robot fleet, strict evaluations, and a rapid path from failure to retraining.

Q8Does this prove Sunday leads robotics?

No. The largest-operation claim comes from Sunday and does not include a public trajectory count that can be compared with every rival. The stronger evidence will be product performance. Sunday says Memo can fold unfamiliar laundry in unfamiliar homes with more than 99% success and plans a home beta this fall. If those results survive wider deployment, the data operation starts looking like a real advantage instead of an internal brag.