Figure launches Helix 2.5 humanoid foundation model that cleans homes with 56% success

Figure announced on September 17 that it has released Helix 2.5, a new foundation model for humanoid robots. The model is described as the most advanced neural network the company has built to date. Helix 2.5 was trained solely on the company's large dataset called Index, which captures human behavior. From this single model the system generated three full-body actions: tidying a living room, folding towels, and making a bed. The tasks involve moving objects, handling rigid and deformable items, coordinating both arms, and active perception. Evaluation was performed in 30 homes without any task-specific data collection, fine‑tuning, or adaptation. The homes used existing furniture such as sofas, beds, and work surfaces that were already present.

Figure claims this is the first demonstration of full‑body zero‑shot generalisation at this scale. The living‑room tidy succeeds only when all scattered toys, ranging from 13 to 15 items, are placed in a basket. Towel folding requires every towel to be folded and placed in a basket. Bed making requires the pillow and duvet corners to be aligned upward and the duvet to be smoothed flat. These criteria were fixed before evaluation. The key metric highlighted is the difference between models that were pre‑trained and those that were not. When comparing a model trained from random initialization with one pre‑trained on Index, the zero‑shot success rate rose from 9% to 56%, a six‑fold improvement. Each evaluated task occupies less than 1.90% of the Index data.

The cost of teaching a single behavior with Helix 2.5 is half that of Helix 02 for comparable actions, expanding the range of applicable homes to 30. Repeatedly doubling the pre‑training data improves downstream robot prediction smoothness and allows loss in the largest training run to be forecasted to four decimal places. Figure says Index generates roughly 35 minutes of human experience each second and has committed computational resources worth 35 billion dollars for Helix training. The company has not disclosed a product launch schedule, price, or target region. It frames the achievement as the first evidence that general humanoid intelligence can learn from human experience and transfer it to new situations.

Reported by BRIDGE.