Breaking

Figure Says Helix 2.5 Generalized Household Tasks Across 30 Unseen Homes

September 17, 2026 at 16:37 EDT

Robotics
Foundation Models

Figure introduced Helix 2.5 on September 17, saying its humanoid robot completed household tasks across 30 Bay Area homes it had never seen, without collecting data, fine-tuning, or adapting in those environments. The company framed the result as a step toward home robots that can transfer learned physical behavior from one setting to another instead of being trained house by house.

September 17, 2026 · Figure

Helix 2.5 takes the home test

Figure says one humanoid policy handled tidying, towel folding, and bed making across unfamiliar homes without environment-specific retraining.

The generalization gap

9%Trained from scratch

56%Index-pretrained

A comparable zero-shot evaluation rose by more than 6x when the policy started from Index pretraining.

30Unseen homes

3Whole-body behaviors

8xPretraining data range

0.54%Forecasting error share

How the trial was framed

Index pretraining
Human behavior data builds a broad physical prior.

→

Behavior adaptation
Tidying, towel folding, and bed making are specified.

→

Zero-shot homes
No evaluation-home data, fine-tuning, or adaptation.

Why it matters

The result suggests household behavior data may transfer across new layouts, furniture, textiles, and objects.

What remains open

A 56% success rate and three evaluated behaviors leave reliability, breadth, and home safety as the next hurdles.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.