Figure · 2025
Helix: a VLA that runs entirely on the robot
Helix is Figure's in-house vision-language-action model for its humanoids. Its signature is whole upper-body control — arms, wrists, fingers, torso, head — driven by a dual-system model that runs fully on-board, no cloud round-trip.
What it is
Slow understanding, fast hands
Helix splits into two systems that run together. System 2 is an onboard vision-language model that understands the scene and the spoken instruction at a few hertz. System 1 is a fast visuomotor policy that turns that understanding into high-rate motor commands across the whole upper body.
It is the same dual-system pattern that defines the modern VLA architecture — Helix's distinguishing bet is running all of it on embedded compute, which is what real-time, safe humanoid control demands.
Why it matters
On-board is the point
Motor control that keeps a humanoid safe cannot wait on the network. By running Helix fully on-device, Figure ties latency, reliability, and safety to hardware the robot carries — a design choice that mirrors where the whole field is heading.
Helix gives the robot capable hands. What it does not give is memory of your home or worksite — the knowledge layer that benned builds so a capable robot also arrives already knowing your world.
FAQ
Common questions
What is Helix?
Figure’s in-house vision-language-action model for humanoid robots (2025). It uses a dual-system design — a slower onboard VLM for understanding plus a fast visuomotor policy for high-rate control — and runs entirely on-board.
What makes Helix different?
It was presented as controlling the entire humanoid upper body — arms, wrists, fingers, torso, head — at high rate while running fully on embedded onboard compute, with demos of multiple robots collaborating on one policy.
Last updated: July 2026