How Weave Annotates Complex Edge-Cases for Robots Folding 1,000 lbs of Laundry per Week

Introducing Customer: Weave Robotics

Weave Robotics designs and builds home robots. Their first generation robot, Isaac 0, has already been deployed in hundreds of homes and businesses, where it has been folding >1,000 lbs of laundry each week.

With almost a year of live deployment data and learnings, the team are now working on Isaac I: a wheeled robot with a friendly exterior, soft-fabric covering, two arms and two-finger parallel grippers, designed to take on broader chores from tidying rooms and putting away toys to making beds. Isaac runs two proprietary models in production: a vision language action (VLA) model and a progress model that supervises task completion in real time.

The Problem: Internal tooling couldn't keep pace with evolving annotation needs

Weave’s models require richly annotated training data for an environment that is, by definition, always different. No two homes are the same, and even the same home undergoes daily changes. That variability creates a core challenge for model performance, and persists as Isaac robots are deployed in more environments.

Before Encord, the team at Weave relied primarily on open source tools. As Weave co-founder, Evan Wineland, explains, the tools were not sufficiently turnkey for the annotation tasks that the team needed. For a team building robots that encounter novel real-world scenarios every time a new unit ships, that gap had direct consequences.

The team wanted to close three main gaps: 1) Getting access to more information on how annotation was progressing 2) Having a mechanism to catch issues, and drive improvements in quality, and 3) Being able to iterate on the task definition itself.

As a result, the annotation process couldn’t keep pace with the expanded variety of data the robots were encountering, and the team wanted to ensure the models would not fall behind the real world.

The Solution: Tools and analytics that evolve alongside the data

The Weave team came to Encord looking for two things: a consistent, high-quality throughput on annotation tasks that were ever-evolving, and a platform that was easy to use.

"We need to be able to quickly dig into our own data using Encord's platform, to understand when our process is breaking and evolve our annotation and curation pipelines over time. This becomes more and more important as we deliver more robots."

The team used Encord to quickly set out an initial goal, and then adjust as their understanding of the task deepened. Weekly syncs with the Encord team made that iteration structured and fast. As their co-founder described it: "With Encord, we're able to rapidly iterate on what we’re looking for, and make progress fast.”

The analytics tools proved equally important as a way to surface quality metrics they hadn't previously defined. "What Encord allows us to do is to find additional quality metrics that we maybe hadn't previously conceived of before," their co-founder noted, "and revise our own pipeline so that we can refine the annotation task itself."

Results: Autonomous rollouts at production scale, with a data process that scales with them

Weave's first autonomous rollouts were underpinned by data annotated with Encord. The robots don't run with remote supervision; they operate independently across homes and commercial laundry environments. This required training data that accurately reflected the variability of those environments, annotated to a level of granularity that the team itself defined and refined over time.

"Encord made a bet on us when we were an even younger and even less established company. That made a huge difference for the quality of our data and for our very first autonomous rollouts, which have been a key part of our success."

As Weave rapidly expands to more environments and new products, the value of a robust data pipeline compounds. Every day in deployment brings new edge cases, new scenarios that weren't in the training set. "The real advantage of having a partner like Encord is that we don't have to worry about growing into new tasks with them."