Helix is Figure’s proprietary on-board Vision-Language-Action model, first announced in February 2025 and significantly upgraded as Helix 02 in January 2026. The original Helix introduced a dual-system architecture inspired by Kahneman’s ‘Thinking, Fast and Slow’ framework: System 2 (S2) is a roughly 7-billion-parameter open-source VLM that handles slow, semantic reasoning over scenes and natural-language goals, while System 1 (S1) is a much smaller (~80M-parameter) transformer that converts S2’s latent representations into 200 Hz visuomotor control signals. This split lets a single set of neural-network weights drive 35 degrees of freedom across torso, head, wrists, and individual fingers, including coordination between two robots. Helix 02 extends the architecture to a hierarchical ‘System 0/1/2’ design covering whole-body locomotion, balance, and manipulation in a single learned model, demonstrated by an unedited 4-minute video of a Figure 03 humanoid autonomously unloading and loading a dishwasher across 61 sequential actions. Helix is purpose-built for the Figure humanoid product line and is not available outside Figure, but it represents one of the most capable production-deployed humanoid AI systems currently disclosed.
End-to-end vision-language-action model from Figure that runs entirely on the Figure 02 and Figure 03 humanoids. A dual-system architecture (System 2 ~7B VLM for reasoning, System 1 ~80M transformer for fast control) enables 35-DoF whole-upper-body control. Helix 02 (Jan 2026) extends to whole-body autonomy.
Helix is Figure’s proprietary on-board Vision-Language-Action model, first announced in February 2025 and significantly upgraded as Helix 02 in January 2026. The original Helix introduced a dual-system architecture inspired by Kahneman’s ‘Thinking, Fast and Slow’ framework: System 2 (S2) is a roughly 7-billion-parameter open-source VLM that handles slow, semantic reasoning over scenes and natural-language goals, while System 1 (S1) is a much smaller (~80M-parameter) transformer that converts S2’s latent representations into 200 Hz visuomotor control signals. This split lets a single set of neural-network weights drive 35 degrees of freedom across torso, head, wrists, and individual fingers, including coordination between two robots. Helix 02 extends the architecture to a hierarchical ‘System 0/1/2’ design covering whole-body locomotion, balance, and manipulation in a single learned model, demonstrated by an unedited 4-minute video of a Figure 03 humanoid autonomously unloading and loading a dishwasher across 61 sequential actions. Helix is purpose-built for the Figure humanoid product line and is not available outside Figure, but it represents one of the most capable production-deployed humanoid AI systems currently disclosed.