Figure Trains Helix 2.5 to Clean 30 Unseen Homes

Figure has announced Helix 2.5, a humanoid robot policy that successfully completed household tasks in 30 rented Bay Area homes that were entirely excluded from its training data. Operating under a zero-shot framework with no fine-tuning or local adaptation, the robot performed long-horizon tasks including tidying living rooms, folding towels, and making beds. The system treats these varied domestic environments as whole-body control problems, coordinating locomotion, active perception, and bimanual manipulation to recover from physical errors.

To isolate the impact of pretraining, Figure compared a model trained from scratch against one initialized with Index, its massive library of human-behavior video. In blind evaluations, the policy trained from scratch achieved a mere 9% success rate, whereas the Index-pretrained Helix 2.5 reached a 56% success rate. This represents a 47-percentage-point increase, or roughly a 6.2-fold improvement. Furthermore, Helix 2.5 matched the performance of the older Helix 02 model while requiring only half of the task-specific adaptation data, despite being tested across 30 times as many environments.

The company also demonstrated what it calls the first human-to-humanoid scaling law. By training four models across an eightfold range of Index data, Figure mapped a highly predictable decline in action-prediction loss. This relationship allowed engineers to forecast the final test loss of the largest training run to four decimal places, with a forecast error of just 0.54% of the measured loss variation. To support this trajectory, Figure has committed $3.5 billion in compute to the Helix project, with its Index dataset currently expanding by approximately 35 minutes of human video every second.

For robotics practitioners, these findings offer a concrete blueprint for deploying humanoid systems in unstructured environments. The results suggest that pretraining on diverse human video, followed by targeted adaptation on smaller robot datasets, is a highly viable path to achieving zero-shot generalization. While Helix 2.5 still failed 44% of its pooled trials and has not yet been released publicly, the predictable scaling of its action-prediction loss provides a reliable framework for teams to estimate performance gains before investing in expensive training runs.