Sunday Robotics achieves 99.1% success folding laundry across 785 zero-shot attempts, introduces 'solve' framework for robotics benchmarks

Sunday Robotics is building Memo, a home robot focused on household tasks, and its latest result is the kind of number that reframes what's possible in physical AI: 99.1% success folding laundry across 785 zero-shot autonomous attempts, across diverse garments and unseen environments.

The 99.1% figure matters precisely because of how Sunday defines it. Zhao introduces a framework he calls "solve" to cut through what has become a credibility problem in robotics benchmarks. A success rate only means something, he argues, when you've locked down two variables: scope (what range of situations the robot must handle) and adaptation budget (whether the model is allowed to train on deployment conditions before being tested). Sunday's 785-attempt result was zero-shot, meaning the model saw no examples from the test environments beforehand. Zhao draws the contrast to self-driving: 99% on a closed course and 99% on city streets are not the same number.

“After filling close to a thousand garments, the success rate is ninety nine point one percent, which is very, very reliable... It turns out that when we scale up pretraining, you can do one shot learning with these models — you can teach the robot to fold your shirt in a new way with one demonstration, and the model can extrapolate that to an unseen shirt on an unseen bed.”

One-shot personalization

The more surprising finding from the training run is what emerged from scaling pretraining compute and data: one-shot learning. The robot can now learn to fold a shirt in a new way from a single demonstration and generalize that to an unseen shirt on an unseen bed. Zhao says this wasn't engineered in directly — it emerged from scale. The practical implication is user personalization: teaching Memo where to put clothes, how to organize a living room, or how to handle specific garments via a single video or photo. Zhao puts that as a shippable feature in late 2025 or early 2026, not this year.

Task list and the path to physical AGI

Zhao draws a distinction between two objectives Sunday is running simultaneously. The first is the minimum viable task list for a home robot to justify its existence — he thinks that list is short, because genuinely annoying household chores are few and repetitive. Laundry, dishwasher loading, and a handful of others may be enough. The second objective is what he calls "research market fit": the research required to make Memo useful at home is the same research advancing toward generalized physical intelligence. Because Sunday's training recipe makes no hardware-specific assumptions, the same dataset and model that runs Memo today could, in principle, run a legged robot in the future.

On cooking — the obvious next milestone — Zhao is candid that it's riskier than laundry. A botched fold leaves a wrinkled shirt. A botched cooking attempt leaves broken glass and olive oil on the floor. Laundry, he argues, is actually the harder research problem precisely because the consequences of failure are low enough to generate clean training signal at scale.

The capability roadmap, as Zhao describes it, is straightforwardly exponential: no task-specific assumptions in the training recipe means Sunday could pursue a thousand manipulation skills in parallel, limited only by data and compute.