Figure Says Helix 2.5 Took a Humanoid Into 30 Unfamiliar Homes With No Setup

A person walks into a house they have never visited and starts tidying without being shown where anything goes. Figure says that is still not how robots work, because they learn the places they operate one place at a time. In a post on its own news page, the company introduced Helix 2.5, which it calls the most advanced neural network it has built, and framed it as an attempt to answer a harder question: can a humanoid enter an unseen home and immediately get to work, with its whole body, on its own.

Figure says the previous generation, Helix 02, had already shown that a neural policy could coordinate a humanoid's whole body over long stretches of time, from unloading a dishwasher to running a logistics task autonomously for 200 hours. The catch, according to the company, was that those systems learned from data collected where the robots would operate. Helix 2.5 was pretrained instead on Index, which Figure describes as its global-scale dataset of human behavior, and from that single foundation model the company produced three behaviors: tidying living rooms, folding towels and making beds.

The evaluation is the part Figure leans on hardest. The company says it took the robot into 30 Bay Area homes with zero data collected in any of them, with no fine-tuning or adaptation in those environments or on the objects being manipulated, and that each task used a single fixed checkpoint across all 30 homes. It says success required completing the entire task, with no partial credit: every one of the 13 to 15 scattered toys picked up and placed in a basket, every towel folded and placed in the basket, or both pillows and comforter corners positioned at the top of the bed with the comforter pulled smooth. Figure says that to its knowledge this is the first demonstration of zero-shot whole-body generalization at this scope on a humanoid.

To isolate how much of that came from pretraining rather than from task data, Figure says it trained two policies on identical task-specification data, one from random weights and one initialized from the Index-pretrained model, holding architecture, optimization, hyperparameters and evaluation fixed. In blind evaluations, the company reports the policy trained from scratch succeeded on 9% of zero-shot trials while the Index-pretrained policy succeeded on 56%, more than six times higher. Figure says no single evaluation task makes up more than 1.90% of the Index pretraining dataset.

Figure also says Helix 2.5 matched the success rate of a Helix 02 policy trained to do the same task while using half as much adaptation data, and that the newer model shows a striking qualitative improvement in self-correction, stepping back to reposition, changing stance or walking around a full bed to fix a fold. The company describes that ability to recover and keep making progress as an important effect of Index pretraining.

The post goes further into the training evidence. Figure says it trained four models on nested subsets of Index spanning an 8x increase in pretraining data, and that loss fell predictably with each doubling, which it calls the first human-to-robot transfer scaling law measured on a humanoid. The company says the relationship was precise enough that it could forecast its largest run's test loss to four decimal places before training began, with forecasting error of just 0.54% of the variation across the full data range.

Figure is careful to say the point is not that general humanoid robotics is solved, but that Helix 2.5 is its first evidence that whole-body intelligence can be learned from human experience and transferred to new situations rather than rebuilt each time. The company says Index is now generating roughly 35 minutes of new human experience every second and that it has committed $3.5 billion of compute to training Helix, and it closes by saying it is time to scale up.

Source: Figure (https://www.figure.ai/news/helix-2-5-zero-shot-30-home-generalization).