Figure's Helix 2.5 Helps Robots Learn New Tasks in Strangers' Homes

What happens when you drop a humanoid robot into a home it's never seen before? For Figure, the answer just got a lot more promising.

On September 17, the company released Helix 2.5, a neural network trained on its Index human behavior dataset. To test whether that pre-training actually makes a difference, Figure sent the model into 30 households across the San Francisco Bay Area—homes that had never been part of any data collection. The results? A zero-shot success rate of 56%, compared to just 9% for strategies trained from scratch under identical conditions.

One Model, Three Chores

Helix 2.5 runs on a single Index pre-trained base model and handles three full-body tasks: organizing the living room, folding towels, and making the bed. During evaluation, the homes' layouts, toys, towels, and bedding weren't included in the task specification data. A task only counted as successful if it was completed from start to finish.

Figure also noted that Helix 2.5 matched previous success rates while using only half the adaptive data required by Helix02. And it showed off some impressive self-correction moves—backing up to adjust its position, shifting its stance, even clambering across a bed to reach a stubborn corner.

Isolating the Power of Pre-training

To pin down exactly what pre-training contributes, Figure kept everything else constant: model architecture, optimization method, hyperparameters, downstream data, and evaluation conditions. The only variable was how the model was initialized.

The company says no single evaluation task in the Index dataset accounted for more than 1.90% of the data, meaning the results directly reflect pre-training's effect on generalization. In further experiments, doubling the amount of Index data led to a predictable drop in action prediction loss—a data scaling law for humanoid robots.

The Numbers Behind Index

Index launched on August 25 and has already racked up 264,000 downloads, with over 44,000 weekly active users. Contributors have uploaded more than 16 million videos, earning $15 million in payments. Figure says it has poured $3.5 billion in computing resources into training the Helix series.

For anyone watching the humanoid robot space, those numbers tell a story: data, not just hardware, is becoming the real battleground.

Key Points

Helix 2.5 achieved a 56% zero-shot success rate in 30 unfamiliar Bay Area homes, up from 9% without Index pre-training.
The model handles three full-body tasks—living room organization, towel folding, and bed making—using half the adaptive data of Helix02.
Figure isolated pre-training's contribution by keeping all other variables constant, demonstrating a data scaling law for humanoid robots.
Index has 264,000 downloads, 44,000 weekly active users, and 16 million videos, with $15 million paid to contributors.
Figure has invested $3.5 billion in computing resources for the Helix series.

Figure's Helix 2.5 Helps Robots Learn New Tasks in Strangers' Homes