Helix is a proprietary Vision-Language-Action (VLA) model developed by Figure AI for generalist humanoid control. It unifies perception, language understanding, and motor action in a single model. Figure AI announced Helix on February 20, 2025, under the title "A Vision-Language-Action Model for Generalist Humanoid Control." Successor versions followed: Helix 02, and on September 17, 2026 Helix 2.5, which Figure evaluated zero-shot in 30 homes where no data had been collected.
Architecture
VLA models extend language-model architectures to output not just text but motor actions. In the manner that GPT-4V unifies text and vision modalities, Helix treats actuation as a first-class output modality. Robotics tasks are framed as multimodal generation problems, taking a camera feed and a prompt as input and producing joint commands and reasoning as output, rather than as traditional robotics-controller problems.
Figure AI describes Helix as the first publicly announced VLA model specifically targeting generalist humanoid control.
Later versions
Helix 02
In Figure's account, its second-generation model, Helix 02, showed that a neural policy could coordinate a humanoid's whole body over long horizons, from unloading a dishwasher to running a logistics task autonomously for 200 hours, but its behaviors were learned from data collected where the robots would operate. Helix 02 started from a pretrained vision-language model (Source: figure.ai).
Helix 2.5
Figure introduced Helix 2.5 on September 17, 2026, describing it as the most advanced neural network the company has built. Unlike Helix 02, it was pretrained from random initialization entirely on Index, Figure's crowdsourced dataset of human behavior (see Figure AI). From that single foundation model Figure adapted three whole-body behaviors — tidying living rooms, folding towels and making beds — spanning locomotion, rigid and deformable manipulation, bimanual coordination and active perception (Source: figure.ai).
Evaluation
Figure took the robot into 30 Bay Area homes in which no data had been collected, using each home's existing couches, beds and folding surfaces; no evaluation toy, towel or bedding appeared in the task-specification data, and each task used a single fixed checkpoint across all 30 homes. "Zero-shot" refers to the evaluation environments and manipulated objects, since the tasks themselves were specified through fine-tuning data collected elsewhere. Success required completing the whole task, with no partial credit: all 13–15 scattered toys placed in a basket, all towels folded and placed in a basket, or both pillows and the comforter corners placed at the top of the bed with the comforter pulled smooth. Figure describes this as, to its knowledge, the first demonstration of zero-shot whole-body generalization at this scope on a humanoid (Source: figure.ai).
Contribution of pretraining
In blind evaluations of two policies trained on identical task-specification data and differing only in initialization, the policy trained from scratch succeeded on 9% of zero-shot trials and the Index-pretrained policy on 56%. No single evaluation task makes up more than 1.90% of the Index pretraining dataset. Figure also reports that Helix 2.5 matched the success rate of a Helix 02 policy trained for the same task on data from its evaluation environment while using half as much adaptation data (Source: figure.ai).
Data scaling
Training four models on nested subsets of Index spanning an 8× increase in pretraining data, with model size and downstream training held fixed, Figure reports that held-out robot action-prediction loss fell predictably with each doubling. Using only the smaller runs it forecast the largest run's test loss to four decimal places, a forecasting error of 0.54% of the variation across the data range; Figure calls this, to its knowledge, the first human-to-robot transfer scaling law measured on a humanoid. The result measures data scaling only (Source: figure.ai).
Figure states that the result does not mean general humanoid robotics is solved. It reports that Index was then generating roughly 35 minutes of new human experience every second and that it has committed $3.5 billion of compute to training Helix. All of these figures are the company's own and have not been independently verified (Source: figure.ai).
Related models and approaches
Helix represents an integrated stack in which Figure AI builds both the humanoid hardware and the controlling model. This contrasts with the embodiment-flexible foundation-model approach pursued by Physical Intelligence, whose models are intended to generalize across robot bodies rather than a single integrated platform.
Relationships
- developer: Figure AI
- related: Scaling Laws — Figure reports a human-to-robot data-scaling law for Helix 2.5's Index pretraining
- related: The Industrial Explosion (Davidson, Hadshar), Physical Intelligence (alternative approach), Clone Robotics, Sora and Veo (Video Generation Models) (another multimodal direction), Multimodality, planned AI Robotics.
Sources
- Figure AI / Helix: A Vision-Language-Action Model for Generalist Humanoid Control (2025-02-20)
- Figure AI / Helix 2.5: Zero-Shot 30-Home Generalization (2026-09-17) — figure.ai