Figure launched Helix2.5 humanoid robot neural network on Sept 17. Pre-trained on its Index dataset, the model boosted zero-shot success rates from 9% to 56% in 30 Bay Area households compared to models trained from scratch. Helix2.5 supports full-body actions like organizing living rooms and folding towels, achieving comparable success with half the adaptive data of Helix02. It demonstrates self-correction capabilities and follows scaling laws observed when doubling Index data. Index has amassed 16 million videos and 44,000 weekly active users, with Figure investing $3.5 billion in computing for the Helix series.