AI

Sunday’s ACT-2: Memo Hits 99.1% Laundry Folds in Unseen Homes

Shar Hendrix 5 min read

Sunday Robotics just published the kind of number home-robot demos usually dodge. In its ACT-2 Preview (July 17, 2026), the company says Memo achieved 99.1% ±0.3% success on laundry folding across 785 autonomous attempts in unseen homes — same model checkpoint, zero per-home fine-tuning.

Business Insider interviewed CEO Tony Zhao: Memo folded successfully more than 99% of the time in unfamiliar homes and on garments it was not specifically trained to handle. The company wants that reliability branded as a “Solve” — performance + declared scope + stated adaptation cost — instead of another one-off video.

Sunday Robotics ACT-2 Memo laundry folding preview
Still from Sunday’s ACT-2 Preview film. Source: Sunday Robotics / YouTube.

ACT-2 Preview: Memo folding in homes it has never seen. Source: Sunday Robotics / YouTube.

The Numbers They Put on the Table

From Sunday’s technical blog (self-reported evaluation):

  • 778 successful folds of 785 attempts (99.1%)
  • 9 garment types (T-shirts, pants, shorts, leggings, blouses, etc.; socks/underwear excluded from the fold metric)
  • Mean fold quality 4.72 / 5; 98.3% rated 4★ or better
  • Median completion 2 min 13 sec per successful fold
  • Stress cases they highlight: inside-out pants, lighting extremes, mid-fold disturbances, 8XL vs baby clothes, even a child nearby
Sunday ACT-2 blog hero imagery
Hero art from Sunday’s ACT-2 Preview post. Source: Sunday Robotics.

How ACT-2 Is Trained (Their Recipe)

Sunday says the unlock is a strong pretrained base plus sample-efficient post-training:

  • Pretraining on high-diversity human manipulation data from Skill Capture Gloves (~$200 to make; BI notes glove form matches Memo’s hands)
  • Memory Developers collecting “memories” across many homes (humanoid.guide cites 2,000+ glove users / ~500 homes in secondary coverage)
  • Claim: a single fine-tuning example can teach a new behavior that generalizes to unseen environments
  • In-house fleet hill-climbs edge cases; gains supposedly transfer “to the wild”

Memo is wheeled, with a telescoping spine — not a biped. BI reports Sunday raised $165M earlier this year at a $1.15B valuation. Company site: beta late 2026; hand-built units cost ~$20k today, retail TBD after beta.

A Human’s Take

I love that they published error bars and a garment table. That’s still self-graded homework — but it’s more honest than a silent success reel. Fall beta is when the claim meets other people’s laundry piles. If those numbers hold outside Sunday-selected homes, the wheeled home-robot lane just got a real yardstick.

Sources