Our plan for autonomously folding your laundry

At Weave we're building capable, general-purpose, and safe home robots that help you get more done. One of the first things Isaac will help our customers with is folding laundry, and to that end we want to share a bit on the progress we've made.

Here we show in one shot how much of the end-to-end folding workload Isaac can handle for customers. Isaac is able to handle back-to-back t-shirt folding and any necessary workspace clearing, thanks to our blend of autonomy with teleoperation and the mobility of our custom hardware. This early prototype Isaac folds shirts end-to-end 70% autonomously, with a human helping only as needed.

Not only is our approach key for collecting valuable, diverse data for further improving our own models (more on this below); it also makes for the best product by guaranteeing that a general-purpose robot actually folds clothes end-to-end, with quality. And most importantly, this means we can start shipping Isaacs to customers soon, which is the best possible way to rapidly improve the product: by learning from experience in real homes, doing real work.

Thoughts on successful laundry folding

Folding clothes not for a demo but as part of a shipping product comes with a certain bar, starting with how clothes are folded and how they end up.

We think folded clothes should look fairly clean. To most people, in the case of shirts that means they look reasonably uniform and have few or no stray corners. They should face the same way when stacked, and if there’s a collar, it should probably face up. How Isaac folds also shouldn’t make too many assumptions about its surface: folding steps shouldn’t need a table; ideally they work on multiple reasonable surfaces (beds, for example).

We also assume Isaac will need to be able to move clothes from one place to another. This is in part because most of Weave’s preorders are from families (which generate lots of clothes), which means Isaac needs to be able to continuously clear its workspace, like it does above.

There are lots of steps to folding clothes that it would be convenient not to do. We also still plan to improve how shirts look when Isaac’s done folding. But, we think making a real product implies something like the above, even for a first-of-its-kind.

How do we get there from here?

A shipping product (rightfully) comes with the expectation that core functionality works most of the time, where what “works” means is not conveniently defined. Yet the constantly changing nature of the home is a key part of what has limited adoption even of limited-purpose robots that try to navigate homes autonomously. At the same time, deploying robots in new environments is the best way to collect diverse data and in turn make robots more autonomously capable and robust.

Our approach solves this apparent chicken-and-egg problem while making sure that Isaac gets the job done. We start by prioritizing a baseline of autonomous capability by training our own vision-language-action (VLA) model with data we collect. We’ve also built a performant networking stack so that Isaac can be remotely assisted by a human operator when needed, who ensures that Isaac completes its task. And finally, we’ve built a data pipeline that allows each new VLA to be trained with (learn from) previous task completions. This makes Isaac capable of learning and improving from real-world experience as well as in-house data collection.

We’re already collecting data this way every day with our fleet across San Francisco. As we build and deploy more robots, our dataset will only get more diverse and valuable—making our models more robust and faster.

In closing

That may have been more than you expected to think about laundry this week.  We’re thinking about laundry a lot so that customers won’t have to!  There’s a lot more we’ll share soon: other articles of clothing, speeding up, refining our custom hardware, and modeling and inference improvements.

The steps we’ve taken (how we fold, Isaac’s design & form factor, our data collection & inference stack) fortunately mean we have line of sight on everything we’ll need to ship delightful, general-purpose robots that actually fold laundry in customer homes. By shortening the timeline to a capable and useful product, we’ve set ourselves up to quickly iterate toward the best product: a general-purpose robot that continually learns how to be useful in the real world.