Helix 2.5: Zero-Shot 30-Home Generalization

This measures data scaling only, with model size and downstream training fixed. But it suggests that a property central to language-model scaling may extend to humanoid robotics: estimating the benefit of the next doubling of human experience before training on it.

It’s Time to Scale Up

A year ago, we formed the thesis that robotics was off by multiple orders of magnitude: in model size, compute, and, especially, data. Index was our attempt to close that gap by asking a simple question: can a humanoid learn enough from human experience that it no longer has to be taught every place it will eventually work? Behavior specification got 2x cheaper while its scope expanded 30x. Helix 2.5 used half as much task-specific data as a representative Helix 02 behavior, then generalized that behavior across 30 unseen homes. A human-to-humanoid robot transfer scaling law. Repeatedly doubling Index pretraining data improved downstream robot-action prediction smoothly enough to forecast our largest run’s loss to four decimal places before training. Video 6: Whole-body error recovery: the robot steps back, repositions, and moves around the environment to recover from mistakes and complete long-horizon tasks.

A Human-to-Humanoid Robot Transfer Scaling Law

Figure 2: A Human-to-Humanoid Robot Transfer Scaling Law: Forecasting error on the largest run was just 0.54% of the variation across the full 8× data range.

Scaling laws transformed language model development by showing that next-token prediction improves predictably with more data and compute, allowing large runs to be forecast from smaller ones.