Sunday’s ACT-2 Claims 99.1% Laundry Success Before Its First Home Beta

• Sunday reports 778 successful folds in 785 autonomous attempts, using a fixed model and no post-training or expert demonstrations in each evaluation home.
• The 99.1% result covers nine specified garment types and excludes socks, bras, underwear and accessories; it is not an independent test of a general-purpose home robot.
• Memo’s price and beta size remain undisclosed, while rival home robots have announced prices and human-assisted operating models.

Sunday Robotics has put an unusually detailed number behind one household task. The result makes ACT-2’s transfer across homes worth testing, but it does not yet show whether Memo can deliver enough autonomous work at a support cost households will accept.

What the 99.1% result measures

Sunday said in its ACT-2 preview that its wheeled Memo robot folded and stacked a garment in 778 of 785 autonomous attempts. That is a 99.1% success rate, with a company-reported standard error of 0.3%.

The evaluation covered T-shirts, thick and thin long-sleeved tops, polos, sleeveless tops, blouses, pants, leggings and shorts, in sizes from XXS to 8XL. Items began naturally crumpled in piles or baskets on beds or the floor. Memo operated from different positions in unseen rooms with varied surfaces, lighting and sheet colors.

The boundary excludes socks, bras, underwear and accessories. Sunday also does not disclose how many homes were used. The physical evaluation garments and data from the evaluation homes were not used for task-specific post-training or model selection, the company says, and one fixed checkpoint and system configuration were used throughout.

Sunday’s own measurements add context to the headline rate:

The numbers come from Sunday, not an independent replication. They establish performance within the company’s declared test, not reliability across all laundry or other household work.

Sunday calls reliable performance across a declared scope and at a stated adaptation cost a “Solve.” A report reproducing the company’s statement captured the distinction Sunday wants the label to make:

“A demo asks if a robot can do something once. A solve asks if a robot can do something reliably while the world changes around it.”

That is a useful reporting framework, but “Solve” is Sunday’s proposed standard. The evidence here is still a company-designed evaluation of the company’s model on the company’s robot.

ACT-2’s claim is about the improvement loop

Laundry folding itself is not ACT-2’s main technical claim. Sunday argues that broad pretraining on its proprietary sensorized-human dataset makes later post-training on a small amount of curated robot data transfer to homes outside its facilities.

In one internal scaling experiment, Sunday applied the same post-training procedure at several pretraining-data levels and ran 50 evaluations behind each reported success rate. The company says the gap between in-domain and out-of-domain folding success fell from 82 percentage points with no pretraining data to zero at the full scale tested. The result supports Sunday’s proposed mechanism inside its own setup; it does not show that every new task will transfer equally well.

A separate one-example exercise was smaller still. Sunday post-trained four copies of the same base model, giving each one demonstration of a different folding technique. Each copy then reproduced its assigned technique on one held-out garment. Four successes across four techniques show that those behaviors transferred in that exercise. They are not the basis for the 785-attempt reliability result, which followed broader pretraining and iterative post-training.

The potential business leverage lies in controlling the entire loop. Sunday owns Memo’s hardware, the model, an in-house robot fleet and its data operation. Sensor-equipped gloves that mirror Memo’s hands cost about $200 to make, according to a report based on company interviews. When the in-house fleet exposes a failure, Sunday’s thesis is that a targeted fix can improve a fixed model before it reaches many homes, without gathering a fresh expert demonstration in each one.

That infrastructure requires money before it produces a consumer business. Founded by Tony Zhao and Cheng Chi in 2024, Sunday has grown to more than 100 employees. It raised $165 million earlier in 2026 at a $1.15 billion valuation. The financing gives the company capacity to build its fleet and data operation; it does not answer Memo’s manufacturing cost, service cost or selling price.

Prior systems and rivals narrow the distinction

Robotic laundry folding predates ACT-2. The 2022 SpeedFolding research system learned from 4,300 human-annotated and self-supervised actions, then reported a 93% success rate and an average completion time under 120 seconds on unseen garments. The published abstract says it generalized across color, shape and stiffness.

Those figures do not create a direct leaderboard. SpeedFolding was a bimanual system following user-defined folding-line instructions; Sunday’s measure includes retrieval, autonomous retries, stacking and transfer across its declared home distribution. ACT-2’s reported mean of 2 minutes 19 seconds is longer, but it claims a broader operating setup and zero adaptation per home. Different bodies, instructions and success criteria prevent either percentage from proving overall superiority.

Commercial alternatives also redistribute Sunday’s advantage. A report on Memo says Weave Robotics plans to start California deliveries of its $7,999 Isaac 1 in the fall. 1X says in its product announcement that NEO can be bought through a $20,000 early-access option for delivery in 2026, with a $499 monthly subscription planned for later.

The operating models differ as much as the prices. For an unknown chore, 1X says an owner can schedule a company expert to guide NEO while completing the task and helping it learn. Zhao says Memo will act autonomously, with a remote operator stepping in only when a customer needs help, and that Sunday does not plan to use those interventions as training data from customer homes.

Sunday’s model may reduce the amount of customer-home data in its improvement loop. It does not eliminate remote assistance, and no disclosed figure shows how often Memo will need it. Intervention frequency therefore connects the privacy claim to an unresolved cost: frequent exceptions could still make support expensive even if the robot is autonomous by default.

The beta must reveal intervention, price and a second reliable chore

Sunday plans to place Memo with families this fall after testing in employees’ homes and short-term rentals, according to the report. Zhao declined to disclose the number of beta homes, and none of the retained material gives Memo a price.

The company is working on vacuuming, toy organization, zipper fastening, turning pants inside out and coffee preparation. Sunday says none has yet been tested against its Solve standard. An earlier model also demonstrated clearing a table, loading a dishwasher and pulling an espresso shot, but a demonstration is precisely the evidence threshold ACT-2 is meant to supersede.

The beta can resolve the central question only if Sunday reports operational evidence that its folding test does not contain: intervention frequency, uptime, safety incidents, support cost, task frequency and customer willingness to pay. A second chore reaching the same declared reliability threshold would also show whether ACT-2’s post-training advantage extends beyond variations of one task.

Until those results arrive, the defensible conclusion is narrow. Sunday has reported a strong laundry-folding result with unusually explicit scope and adaptation rules. Whether that becomes a useful home-robot service depends on the cost and reliability of the rest of the system, not another folding percentage.