Figure Releases Helix 2.5: Robot Completes Household Tasks in 30 Unfamiliar Homes with Zero-Shot Learning

Humanoid robotics company Figure has unveiled its next-generation neural network, Helix 2.5. The company introduced the robot to 30 previously unseen households in the San Francisco Bay Area, allowing it to autonomously complete three long-duration, full-body tasks—organizing the living room, folding towels, and making beds—without collecting local data, fine-tuning, or adapting to the environment. Figure stated that Helix 2.5 is pre-trained on its human behavior dataset, Index, and achieved a zero-shot full-task success rate of 56% in blind tests, compared to 9% for models trained from scratch, while requiring only half the task-specific data of the representative Helix 02 tasks. Figure also revealed that Index currently adds approximately 35 minutes of human behavior data per second, and the company has committed $3.5 billion in computational power for Helix training.