Figure, a California-based humanoid robot company, announced on September 17th, 2026, that its latest neural network model, Helix 2.5, enabled its Figure 03 robot to perform household tasks in completely unfamiliar homes without any site-specific data collection or fine-tuning. The company rented 30 residences across the San Francisco Bay Area that the robot had never entered and tasked it with tidying living rooms, folding towels, and making beds. According to Figure, the robot achieved a zero-shot success rate of 56.00 percent on complete tasks, compared with a 9.00 percent success rate for a baseline model trained from scratch under identical conditions. This result, described by Figure as a first-of-its-kind demonstration of zero-shot whole-body generalization in humanoid robotics, signals a meaningful step toward robots that can adapt to new environments rather than requiring per-home training.
The technical foundation of this breakthrough is Index, a large-scale human behavior dataset that Figure has been building since early 2026. The platform crowdsources videos of people performing everyday tasks in real environments, generating roughly 35.00 minutes of new human experience data per second. In a controlled experiment, Figure trained two models with identical architectures, task-specific data, and optimization settings, varying only whether Index pre-training was used. The Index-pre-trained model achieved 56.00 percent zero-shot success, while the model without pre-training succeeded in only 9.00 percent of trials, a more than sixfold improvement. Notably, no single evaluation task accounted for more than 1.90 percent of the Index pre-training data, suggesting the gains stem from broad physical-world understanding rather than memorization of specific chores.
The evaluation criteria were deliberately stringent. For the living room task, the robot had to collect all 13.00 to 15.00 scattered toys into a basket. For towel folding, the robot needed to complete the fold and place the towel into a basket. For bed making, pillows and comforter corners had to be positioned within the top one-third of the bed and smoothed. Each task required end-to-end completion, with no partial credit awarded, and any human safety intervention automatically counted as a failure. The robot also demonstrated self-correction behaviors, such as stepping back to reposition, changing posture, or walking around the bed to adjust its approach – capabilities that Figure attributes to the Index pre-training.
Despite the progress, significant limitations remain. The 56.00 percent success rate means the robot still failed nearly half of its complete trials, with toy tidying proving particularly difficult at a 40.00 percent success rate. Figure itself stated that the result does not mean general humanoid robotics is solved. The company has committed 3.50 billion U.S. dollars in compute resources to train Helix and plans to expand Index data collection by 100.00 times over the next 12 months. The key question for the industry is whether scaling data and compute can transform this partial generalization into the reliability required for unsupervised household deployment – a threshold that current results approach but have not yet reached.