Figure AI unveils Helix 2.5 AI model for humanoid robots

Figure AI has unveiled Helix 2.5, a new AI control model for its Figure 03 humanoid robots designed to improve task success through zero-shot generalization. In tests conducted across 30 previously unseen homes in the Bay Area, the model achieved a 56% success rate across three household chores: making beds, folding towels, and tidying living rooms.

This performance represents a significant increase over a baseline model trained from scratch, which achieved only a 9% success rate. The high success rate is attributed to the Index data platform, a pretraining dataset consisting of over 16 million video segments of human behavior.

Specific task performance included a 67% success rate for making beds, 62% for folding towels, and 40% for tidying toys. The testing emphasized the robot's ability to reason about unfamiliar objects and spatial configurations without specific prior training for those environments.