Ask a robot to "tidy the living room" and you're really asking it to make dozens of small judgment calls: what counts as clutter, what's fragile, what shouldn't be touched at all, and what order to even do things in. None of that is written down anywhere. It has to be inferred.

Turning Vague Requests Into Clear Actions

Our models break broad instructions down into smaller, testable steps, checking context at each stage rather than acting on a single rigid command. This lets a robot pause mid-task if something doesn't match expectations, instead of barreling ahead on a plan that no longer makes sense.

Learning From Everyday Correction

When a person nudges the robot toward a different choice, that correction feeds directly back into how future requests are interpreted. Over time, the system builds a working understanding of a household's specific preferences, not just general rules about tidiness or organization.

“A good home AI isn't the one with the most features. It's the one that understands what you actually meant.”

That's the gap we spend most of our research time closing, one household task at a time.

This is also why we're cautious about overpromising general intelligence. A robot doesn't need to understand everything about a home. It needs to understand its own home, deeply and specifically, which turns out to be a much more useful goal to design toward.