This Robot Folded 785 Loads of Laundry. The Real Breakthrough Was Doing It in Unseen Homes

Folding a T-shirt once makes a good robot demonstration. Folding hundreds of unfamiliar garments in homes the robot has never entered is a much harder test.

Sunday Robotics says its wheeled home robot, Memo, completed 778 of 785 autonomous laundry-folding attempts using a new artificial intelligence model called ACT-2. That translates to a reported success rate of 99.1%, with no training or adjustment for each home.

Why laundry is a serious robotics problem

Clothes are deformable objects. Their shape changes every time they are dropped, lifted or folded, and fabric can hide edges, twist sleeves and collapse under a robot’s grip. A machine cannot rely on the fixed coordinates that work on an assembly line.

The home adds another layer of difficulty. Beds differ in height and color, lighting changes, garments arrive crumpled in baskets or piles, and the robot may approach from different positions. A system that performs well in its developers’ laboratory can fail when any of those details change.

What ACT-2 claims to change

Sunday says ACT-2 combines broad pretraining with a rapid post-training process. Human demonstrators first perform tasks while wearing sensor-equipped gloves designed to map their hand movements onto Memo’s robotic hands. The robot fleet then practices the behavior and learns from failures in controlled settings.

The important claim is not that Memo learned to fold clothing. Robots have done that before. It is that improvements made inside Sunday’s own facilities transferred to unseen homes and unfamiliar garments without collecting new task-specific data there.

The company calls this “zero-shot” deployment. In this context, it means the same model checkpoint and system configuration were used without fine-tuning for a particular home or item of clothing.

What the 785 attempts actually covered

Sunday reports that ACT-2 was tested across nine main garment types, from T-shirts and trousers to leggings and blouses. The starting conditions included clothes placed in baskets, piles, on beds and on the floor. Memo operated from different sides of the bed and under varied lighting and surface colors.

Of the 785 attempts, 778 ended with an autonomously folded and stacked garment. Blouses were the most difficult category, with a reported 94.7% success rate, while several easier categories reached 100%. Completed folds received an average quality score of 4.72 out of five under the company’s rubric.

Some objects were excluded. Socks, underwear, bras and accessories were outside the test because people often pair, sort or hang them rather than fold them in the same way. The evaluation therefore does not mean Memo can handle every laundry item or complete the entire washing process.

A useful challenge to the robot-demo culture

Sunday also proposed a reporting standard it calls a “Solve.” Instead of showing a carefully selected successful video, a company would state the task, environments, objects, number of attempts, permitted intervention and adaptation required at deployment.

That idea addresses a real credibility problem in consumer robotics. Impressive clips rarely reveal how many failed takes occurred, whether a human controlled the machine remotely, or whether the room had been specially prepared. Publishing all 785 trial videos gives outsiders more material to inspect than a conventional highlight reel.

The central limitation

The headline result is still a company-reported benchmark. Sunday designed the test, chose its boundaries, trained the graders and evaluated the outcomes. The company says two annotators reviewed attempts using a rubric fixed before evaluation, but no independent laboratory or peer-reviewed paper has reproduced the 99.1% result.

The tests also measure one constrained skill. They do not establish that Memo can move safely around children or pets, recover from every household interruption, protect private information, operate economically for years or generalize equally well to dishes, cleaning and other chores. Sunday itself says several additional skills are not yet reliable enough.

Why this matters

Artificial intelligence has advanced quickly in language and images because digital data can be copied cheaply. Robots face a different world, where every training example involves motors, cameras, physical objects, time and the risk of damage. Generalizing from a limited training environment is therefore essential if home robots are ever to become useful products rather than expensive demonstrations.

Memo is expected to enter a limited home beta, but Sunday has not announced broad commercial availability. The next proof will not be another polished video. It will be whether ordinary households experience the same reliability, safety and independence claimed in the company’s tests.