Research Reports

Overview

In late February 2025, Figure AI released Helix, a general-purpose (“generalist”) VLA (Vision Language Action) model designed for humanoid robots. This model allows robots to perform actions based on visual identification of objects and environments along with received verbal commands. Helix boasts several innovations, including being the first VLA model to enable rapid and continuous control of a humanoid robot’s upper body, encompassing not only the wrists and fingers but also the torso and head. Furthermore, Helix facilitates cooperative systems among multiple robots. Figure AI’s demonstration video shows two robots working together to organize daily purchased items by picking and passing objects to each other. Helix learns all behaviors through a single neural network, eliminating the need for task-specific fine-tuning. It can grasp most household objects with just verbal commands and is designed to be embedded in low-power GPUs, granting it considerable deployment flexibility and making it already commercially viable to some extent.