Build the Robot, Not the Brain

If you have an idea for an app that needs to understand language, you don't start by training a language model. You call one. The intelligence is already there, behind an API, and you spend your time on the part that's actually yours: the product.

Robotics is not like this, yet.

If you want a robot to do something useful today, you build the brain along with the body. You write your own controllers, assemble data pipelines, and train your own models before you can even start on the task you care about. Many of those pieces aren't just hard engineering; they are open research problems, with no widely known recipe, let alone an off-the-shelf API.

We think that should change. General-purpose robot models can be the shared intelligence that every robot builds on: one model, trained across many tasks and machines, that a team adapts to its own hardware instead of building from scratch. That is how we work at Blink Labs. Our partners build the robot; we provide the brain. When that becomes the norm, the cost of trying a new robot idea drops sharply, and useful robots start showing up everywhere, from homes and hospitals to offices, warehouses, and factories.

The best test of that idea is a partner shipping real robots on it. Below is what changed when Weave built its laundry-folding robots on our models.

Weave builds robots for the home. Earlier commercial versions already work in businesses across the San Francisco Bay Area, doing the same job Weave is bringing to homes: folding laundry.

It is a compelling first task for home robots. Laundry is universal, time-consuming, and attention-intensive, and the workload grows with every person in the household.

It is also one of the hardest manipulation problems in robotics. Folding is a long-horizon task where one misplaced corner can spoil the final fold, and clothing is deformable and endlessly varied in size, fabric, color, and shape. Because Weave folds for laundromats as well as homes, its robots have to handle all of it: not just t-shirts but long sleeves, shorts, pants, and more, in wildly varying sizes and fabrics.

We worked with Weave to push autonomous folding on our models at several locations, including one of their customer sites. Across those deployments, our models improved fold quality, cut the time to fold each garment, and reduced how often Weave's remote specialists had to step in to reach a presentable fold.

Weave now ships to new customers every week, and every deployment feeds the next model. Their data goes into our pre-training, which is where the drop in missed grasps and interventions above comes from.

Weave didn't have to build a robot brain to fold laundry for paying customers. They built the robot, brought their own data, and started from a model that already knew how to handle the physical world. That is the division of labor we want to make normal: hardware teams focused on the machine and the job, and a general-purpose model they can build on for the intelligence. If you are building a robot and want to start there, please get in touch.