Connecting language to a humanoid's movements
Figure develops humanoid hardware and learned control systems. Helix connects visual observations and language instructions to continuous actions, combining a slower vision-language component with a faster policy that responds to the physical scene. This makes Figure a developer of both the machine and the models that drive its behavior. [1]
The original February 2025 Helix research controlled the upper body, including fingers, wrists, torso and head. Figure demonstrated two robots coordinating a manipulation task. That result should not be described as learned control of walking and balance across the entire robot. [1]
Extending control to walking and balance
Helix 02 added whole-body coordination in January 2026. Its hierarchy includes a learned motion prior trained with human motion and reinforcement learning in simulation, beneath the components that interpret goals and generate joint targets. The aim is to coordinate locomotion and manipulation as objects and contacts change. [2]
Figure presented a continuous dishwasher task that it described as autonomous, without intervention or resets. This is a developer-reported demonstration of the control system; it does not establish an independently measured success rate across homes or a generally available household robot. [2]
Sources
- Helix: A Vision-Language-Action Model for Generalist Humanoid Control
Figure · Feb 20, 2025
- Introducing Helix 02: Full-Body Autonomy
Figure · Jan 27, 2026