Figure unveils Helix 2.5 with remarkable zero-shot success rate

Figure released Helix 2.5, a humanoid robot neural network, on September 17, 2026. Pre-trained using the company's Index human behavior dataset, it was tested in 30 Bay Area households with no prior data. Zero-shot success rates rose from 9% to 56%, surpassing scratch-trained models. Helix 2.5 supports organizing the living room, folding towels, and making the bed, achieving comparable results with half the adaptive data of Helix 02 and full-body self-correction. To isolate pre-training effects, Figure kept model architecture, optimization, hyperparameters, downstream data, and evaluation consistent, only changing initialization. No single Index task exceeded 1.90% of the dataset. Action prediction loss decreased predictably with doubled Index data, measuring data scaling law. Index launched on August 25, 2026, with 264,000 downloads, 44,000 weekly active users, and 16 million videos, paying $15 million to contributors. Figure invested $3.5 billion in computing resources for the Helix series.