Figure Releases Helix 2.5 for Autonomous Humanoid Robot Housework in Unseen Homes

Figure AI has released Helix 2.5, a humanoid foundation model pretrained on its global human behavior dataset, Index. Without any prior data collection or environment-specific fine-tuning, robots running Helix 2.5 autonomously performed multi-step household tasks—such as tidying living rooms, folding towels, and making beds—across 30 unfamiliar Bay Area homes. This accomplishment marks a significant advance in zero-shot generalization for embodied AI, proving that whole-body physical skills learned from human data can transfer directly into novel environments. It moves general-purpose humanoid robots much closer to real-world domestic deployment without requiring custom training for each individual home. Pretraining on the Index dataset raised zero-shot task success rates from 8% to 56% while requiring only half the task-specific data compared to the previous Helix 02 neural network. The benchmark evaluated a single foundation model across active perception, whole-body locomotion, bimanual coordination, and manipulation of both rigid and deformable objects.