RoboticsModels Figure

Figure's Helix 2.5 did chores in 30 homes it had never seen

Illustration for the Figure Helix 2.5 zero-shot homes story

Figure rented 30 homes in the Bay Area, kept them out of its training data, and sent a humanoid robot into them to work. No fine-tuning per house, no adaptation, no data collection in those environments.

What was tested

The model is Helix 2.5. Founder and chief executive Brett Adcock posted four hours of footage on September 17, 2026 showing the robot performing tasks across the 30 houses, all of them new to it, with unfamiliar rooms, layouts and objects.

Two numbers describe the change. Specifying a behavior takes half the task-specific data it previously required, and the scope of where a specified behavior works expands from a single environment to 30 unseen ones. Those measure different things and it is worth keeping them apart: one is the cost of teaching the robot a task, the other is how far that task then travels.

Zero-shot is the load-bearing term. It means the robot was not shown these homes, tuned on them, or given a data collection pass in them before being asked to work.

Generalization is the gap that matters

Almost every impressive humanoid demonstration is filmed somewhere the robot already knows. The environment was mapped, the objects were in the training set, the lighting and layout were fixed. Performance under those conditions says relatively little about performance anywhere else, which is why the category has produced years of compelling video and very few deployed machines.

A stranger’s kitchen is the honest test. The counter height is wrong, the objects are in the wrong places, the floor plan does not match, and none of it was collected in advance. Working under those conditions is the requirement that has to hold before these robots reach anyone’s home.

The caveats are the usual ones for a company-run demonstration. Figure selected the homes, and four hours of footage is a selected window. The company does report a success rate: 56% on zero-shot trials against 9% for a policy trained from scratch, counting only fully completed tasks with no partial credit. That is more specific than the genre usually offers, and the 30x scope figure is the number to hold the next demonstration against.

Sources

ANOTHER News is published by ANOTHER, an AI-native content agency. Daily coverage also runs on Instagram.