Most robot demos are staged. The environment is controlled, the objects are pre-selected, and the robot has seen the space before. Figure’s latest system breaks that pattern in a way that’s hard to dismiss. Helix 2.5, the company’s newest neural network for its humanoid platform, announced in September 2026, was tested across 30 Bay Area homes it had never entered, with objects it had never seen, performing tasks that require the whole body to coordinate — and it worked more than half the time.
That’s not a perfect score. But it’s a real one, and it tells you something important about where humanoid robotics is heading.
What Helix 2.5 actually does
The system performs three tasks: tidying a living room by collecting scattered toys into a basket, folding towels, and making a bed. These aren’t arbitrary choices. Each one combines locomotion, object manipulation, bimanual coordination, and active perception. The robot has to move through a space to find objects, reposition its body to reach them, and complete multi-step sequences without being told what to do at each stage.
None of the evaluation homes contributed any training data. None of the toys, towels, or bedding used during testing appeared in the task-specification dataset. The robot walked in cold and used whatever furniture and objects were already there. Success criteria were strict: all toys in the basket, all towels folded, both pillows and the comforter correctly placed on the bed. No partial credit.
The result was a 56% zero-shot success rate across those 30 homes. That figure comes from a model pretrained on Index, Figure’s large-scale dataset of human behavior. A policy trained from scratch on identical task data, without Index pretraining, succeeded just 9% of the time. Same architecture, same downstream data, same evaluation setup. The only variable was pretraining. That gap — 9% versus 56% — is the clearest signal in the whole paper.
Why pretraining on human behavior matters here
The core argument Figure is making is that broad pretraining on human experience, not environment-specific robot data, is what produces generalization. Helix 2.5 was pretrained from random initialization entirely on Index, unlike its predecessor Helix 02, which started from a vision-language model. No single evaluation task made up more than 1.9% of the pretraining dataset, so the system wasn’t quietly overfitting to these specific scenarios.
This mirrors what happened in language models years ago: pretraining on massive general corpora made fine-tuning far more efficient and far more capable than training from scratch on task-specific data. Figure is making a similar bet for physical intelligence. The analogy isn’t perfect, but the experimental results back it up more clearly than most robotics papers do.
There’s also a data efficiency finding worth noting. Helix 2.5 used half as much task-specific data as a comparable Helix 02 behavior, and generalized that behavior across 30 homes instead of one. Less data, much broader coverage. That’s a meaningful shift in how you’d think about deploying these systems.
How this compares to what else is out there
The competitive field here includes Physical Intelligence (pi), whose pi0 model has shown impressive dexterous manipulation, Boston Dynamics with its Atlas platform, and Agility Robotics with Digit. Most of these systems are still operating in controlled or semi-controlled environments. Zero-shot whole-body generalization across dozens of real, unmodified homes is not something any of them have published results on at this scale.
The key claims Figure makes about Helix 2.5 are:
- Zero-shot performance across 30 unseen homes with no fine-tuning or data collection in those environments
- A single foundation model adapted to three distinct whole-body behaviors
- Index pretraining alone raised zero-shot success from 9% to 56%
- Half the task-specific data required compared to Helix 02, with 30x the generalization scope
- A measurable scaling law: doubling Index pretraining data improved downstream performance predictably enough to forecast loss to four decimal places before training
What this means going forward
Figure is careful not to overclaim. The paper explicitly says general humanoid robotics is not solved. But Helix 2.5 is the first published evidence that whole-body intelligence can transfer from human behavioral data to new physical environments at meaningful scale. That’s a different kind of result than most robotics milestones, which tend to celebrate what a robot can do in one place.
For anyone building or funding in this space, the scaling law finding is probably the most important detail. If doubling pretraining data produces predictable, smooth improvements, then this becomes an engineering problem with a clear path forward. That’s a very different situation from one where you’re guessing whether more data will help at all.
Helix 2.5 isn’t a product yet. But the underlying architecture is moving fast, and the gap between lab results and commercial deployment in humanoid robotics is narrowing faster than most people expected two years ago.
