Robots' 'Laundry Folding' Challenge: Why It's the Ultimate Hurdle and Best Practice Problem

At CES this January, LG Electronics unveiled CLOi, which demonstrated putting laundry into a washing machine and folding towels. Figure AI showcased its robots, 'Figure 02' and 'Figure 03,' expertly folding clothes, and Elon Musk's Tesla Optimus also performed a laundry folding demonstration. As such, laundry folding is becoming a common demonstration task for humanoid robots.

Sunday Robotics unveiled its home robot 'Memo' and made clothes folding its first test task. Additionally, Weave Robotics is developing 'Isaac Zero' and 'Isaac One,' robots specialized for clothes folding. In this way, some startups are focusing on developing dedicated robots, featuring laundry folding as a core function.

Unlike rigid objects such as cups or boxes, fabric is a soft material that wrinkles, folds, and overlaps the moment it is touched. Moreover, each item of clothing has different fabrics and sizes, making it difficult for robots to recognize and process them. Manipulating fabric requires understanding its shape through visual information, using bimanual coordination, and delicate force control to avoid damaging the fabric. Furthermore, robots must have the ability to recover from unexpected errors.

Recently, robots have been solving the clothes folding problem using Vision-Language-Action (VLA) models. This model allows robots to learn the entire process from perceiving the state of clothes in a scene viewed by a camera to deciding the movements of hands and arms through data. The key is to enable robots to learn appropriate actions independently in various situations, rather than designing each movement as in the past.

The task of folding laundry is highly efficient for robot training. Laundry items are less likely to be damaged if dropped or folded incorrectly, allowing robots to safely experience trial and error. Furthermore, data collection is easy as the same actions can be repeated indefinitely with a fixed camera, and new starting states can be created simply by crumpling a towel or changing its orientation, making it good for acquiring diverse learning data. Thus, it is considered an optimal task for robot learning due to its low risk, suitability for repetitive learning, and flexible precision requirements.

In 2015, Japan's Seven Dreamers developed 'Laundroid,' a large home appliance that used cameras and robotic arms to recognize and fold clothes. However, it took up to 10 minutes to fold a single item of clothing, and often failed to recognize clothes that were slightly different in shape, leading to its bankruptcy in 2019. Furthermore, robots like 'FoldiMate' required users to manually load clothes and had limitations on the items they could process, hindering commercialization.

Sunday Robotics reported that its robot 'Memo' succeeded in 778 out of 785 laundry folding attempts in an unfamiliar home environment, according to its own evaluation. This represents a high success rate of 99.1%, and the median time taken to fold a single item of clothing was 2 minutes and 13 seconds. These results demonstrate a significant improvement in robot laundry folding performance compared to past failures.