News · 2026-07-17

Sunday Robotics Says Its Robot Folds Laundry Reliably in Homes It Has Never Seen

Sunday Robotics previewed a home-robotics model called ACT-2 on July 16, 2026, reporting 99.1% zero-shot success at folding laundry across diverse homes it had never seen before -- with no per-home setup. The claim is notable less for the task than for the standard: Sunday is trying to move robotics past the cherry-picked demo video toward measurable reliability in random real homes, and it is proposing a new metric to hold itself to.

Key facts

• ACT-2 reported 99.1% success (plus or minus 0.3%) across 785 autonomous attempts with nine garment types, per the company blog.
• Mean fold quality was rated 4.72 out of 5 stars, with a median completion time of 2 minutes 13 seconds.
• Sunday says the model can learn a new generalizable behavior from a single fine-tuning example.

The headline number sits inside a framework Sunday calls a 'Solve': reliable performance across a declared scope at a stated adaptation cost. This is a deliberate rebuke of how robotics results usually get reported. A viral clip of a robot folding one shirt in one lab tells you almost nothing about whether it works in your house, with your lighting, your laundry basket, your oddly shaped blouse. By declaring the scope (diverse unseen homes) and the adaptation cost (zero -- no setup per home) up front, Sunday is trying to make a robotics claim that means something. For ACT-2 laundry, both are pinned down.

The technical story is about generalization. Sunday reports that scaling up pretraining on a large, sensor-rich dataset of humans doing tasks narrows the 'generalization gap' -- the difference between how the robot performs in-house versus in the wild. As pretraining scales, a small amount of curated in-house data becomes highly transferable. The most useful practical consequence is one-example learning: the model can pick up a new, generalizable behavior from a single fine-tuning demonstration. That lets Sunday 'hill-climb' -- find an edge-case failure in the lab, add one example, and watch the fix propagate to homes it has never seen. This is the vision-language-action recipe applied with an unusual focus on reliability rather than raw capability.

The reported details make the result feel real rather than staged. Success varied by garment: shorts and polos hit 100%, while lightweight, floppy blouses were hardest at about 95% because deformable fabric is genuinely difficult to manipulate. The robot also showed unplanned abilities -- retrieving clothing from the floor and recovering after a human interfered with the task mid-fold. Those emergent recoveries are the kind of thing that separates a brittle demo from something that might survive a real household.

Why it matters: robotics has spent years in a 'look what it can do' phase, where impressive one-off demos rarely translated into reliable products. Fei-Fei Li made exactly this critique the same week, noting that almost all robot demos 'have been confined to heavily constrained laboratory setups, with narrow object sets and short task horizons' and none validated at real-world complexity or duration. Sunday's 'Solve' framing is a direct answer: state your scope and your cost, then report reliability, not highlights. Even the choice of laundry is pointed -- it is a high-variability, deformable-object task that has embarrassed robots for years.

The honest caveats: every figure is Sunday's own, from its own evaluation, with no independent replication yet, and a 'preview' is not a shipping product in anyone's home. 'Diverse unseen homes' is Sunday's characterization of its own test set, and reliability on one well-defined task does not imply general home competence. But the reframing is the real contribution here -- if the industry adopts something like the 'Solve' standard, it gets much harder to pass off a lucky demo as a capability, which is exactly the discipline the field's benchmark and evaluation problems have been missing.

Primary source, verified: read the paper →

Ground Truth. (2026, July 17). Sunday Robotics Says Its Robot Folds Laundry Reliably in Homes It Has Never Seen. Ground Truth. https://groundtruth.day/news/sunday-robotics-act-2-laundry-reliability.html

@misc{groundtruth:sunday-robotics-act-2-laundry-reliability,
  title  = {Sunday Robotics Says Its Robot Folds Laundry Reliably in Homes It Has Never Seen},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {jul},
  url    = {https://groundtruth.day/news/sunday-robotics-act-2-laundry-reliability.html}
}

Topics: robotics · embodied-ai · sunday-robotics · reliability · vision-language-action