Figure has introduced the Helix 2.5 model, designed to enable humanoid robots to perform household chores autonomously. The company tested the technology by deploying robots into 30 homes in the San Francisco Bay Area without conducting any prior site-specific training or data collection.
The model was pre-trained on the Index dataset, which contains human behavioral data. This approach allowed the robots to execute three specific tasks: tidying living rooms, folding towels, and making beds. The system integrates motion, rigid and deformable object manipulation, bimanual coordination, and active perception.
Performance metrics show that the Index pre-training significantly improved zero-shot whole-body generalization. The task success rate increased from 8% to 56% compared to previous iterations. Furthermore, Helix 2.5 achieved these results in 30 new environments using only half the task-specific data required by its predecessor, Helix 02.
Figure stated that Helix 2.5 demonstrates that whole-body intelligence can learn from human experience and transfer those skills to new environments without requiring system reconstruction for each location. The project aims to determine if humanoid robots can enter unfamiliar homes and immediately begin working.
