Behind 99% Folding Success Rate: How Far Are Robots From Entering Your Home?

First, look at a set of data: Sunday Robotics claims its ACT-2 model achieves a success rate of over 99% in folding clothes in unfamiliar homes, with test scenarios covering 50 different households and including thousands of various garments. If this figure is true, it means the last mile for robots moving from labs to homes may have already been half-paved.

But as someone who interacts with AI tools daily, my first reaction wasn't excitement, but skepticism. 99% success in unfamiliar homes—how "unfamiliar" is that "unfamiliar"? Do the tested homes have uniform lighting, layouts, and fabric materials? Folding clothes looks simple but actually involves fabric perception, gripping force, folding sequence, and spatial obstacle avoidance; each item is an engineering challenge. My own cross-border e-commerce product selection model claims 95% accuracy, but drops below 60% when encountering niche categories (like pet Hanfu). So for any claimed high success rate, I first ask: How broad is the distribution of your test samples?

However, after carefully reading Sunday Robotics' technical documentation, I admit they did something different. The core of ACT-2 isn't programming all actions through massive data, but using a variant of imitation learning to allow the robot to autonomously extract key action features from human demonstration videos. This means it doesn't need pre-modeling for each home, but makes real-time decisions via visual and tactile sensors. This is similar to how I used GPT-3 to generate copy—not hardcoding templates, but letting the model understand context.

The tech blog mentions that ACT-2 used 2,000 hours of human clothing-folding videos during training, covering different materials, colors, and folding methods. Inside the model, there is an "attention mechanism" specifically focusing on garment edges and wrinkle keypoints.

This explains a key question: Why can it adapt to unfamiliar homes? Because the training data already includes enough environmental variations, such as different lighting, table heights, and clothing placement styles. But note, this is still testing under limited variables. In real homes, you might encounter wet clothes, shirts with buttons, pants pinned down by cats—extreme cases. How much of the 99% success rate covers these?

I understand that Sunday Robotics wants to prove a direction to the market through ACT-2: Home robots don't need to be retrained for every household. This direction is right. But anyone in cross-border e-commerce knows that between "lab success" and "users willing to pay," there are two huge pitfalls: cost and usage barriers.

Regarding cost, the hardware foundation for running ACT-2 is Sunday Robotics' own robot body, with an estimated price above $20,000. This price means it currently can only be sold to geeks and research institutions; ordinary families won't spend this much on a clothes-folding robot. Regarding usage barriers, although the model claims "plug and play," users need to place clothes on the robot's workbench in a specific way before the robot can operate. Compared to the experience of "throwing clothes in the washing machine and pressing a button," it's still levels apart.

So my view has two layers: First, ACT-2 is indeed a technical milestone in generalization capability, proving the feasibility of end-to-end learning in the physical world, which is closer to human learning methods than Boston Dynamics' pre-programmed route. Second, from a commercial landing perspective, it is still missing two key elements to "enter your home": Price dropping below $5,000, and expanding task scope from "only folding clothes" to "folding + tidying + simple cleaning." Otherwise, why would users buy a robot that only folds clothes? They'd rather buy an iron.

Back to trend prediction. I believe the home robot field will enter a phase of "vertical scenario explosion" in the next two years. Models like ACT-2 will be ported to different household chores, such as making beds, setting tables, or even feeding pets. Each task will have a dedicated AI model, but hardware platforms will gradually unify. By around 2028, a "household affairs robot" priced under $5,000 might appear, capable of completing 5 to 8 common chores. But this requires lower sensor costs, breakthroughs in battery life, and more importantly: Users' psychological tolerance for robot errors. After all, a 99% success rate means 1 out of every 100 times might fold a white shirt into a rag, and users usually remember that one failure.

Sunday Robotics' ACT-2 is a good signpost, but not the destination. Just like the AI product selection tool I use, it initially often recommended wrong categories, but with each iteration, my trust in it increased. Robots are the same; they need time.

Original link: https://www.ithome.com/0/978/617.htm