Figure Helix

What Helix is

Figure announced Helix on 20 February 2025, describing it as a generalist vision-language-action model for humanoid control ^[1]. A model of this kind takes in camera images and a text instruction and puts out motor commands. Helix is described as generalist because a single set of neural network weights covers all of its behaviours, without task-specific fine-tuning ^[1].

Two systems running at different speeds

Helix splits the job between two loops that run at very different rates. System 2 is a vision-language model that runs at 7–9 Hz and handles scene understanding and language comprehension ^[1]. System 1 is a fast, reactive visuomotor policy that translates System 2's latent representations into continuous robot actions at 200 Hz ^[1]. The split reflects a general tension in robot learning: understanding a scene and an instruction takes a comparatively heavy model, while keeping a machine balanced and responsive needs a much faster control loop. Helix bridges the two by letting the slow model hand its internal representations to the fast one rather than handing over finished commands.

The two halves are very different in size. System 2 is built on a 7-billion-parameter open-source, open-weight vision-language model, and System 1 is an 80-million-parameter cross-attention encoder-decoder transformer ^[1].

The action space

Helix coordinates a 35-degree-of-freedom action space covering individual fingers, wrists, end-effector trajectories, head gaze and torso posture ^[1]. That range matters for a humanoid: a single model has to move hands, arms, head and body together, rather than driving one limb at a time.

How it was trained

Figure trained Helix on about 500 hours of teleoperated multi-robot, multi-operator data ^[1]. Because that data was recorded by different people on different machines, the company used an auto-labeling vision-language model to generate hindsight instructions from segmented robot video clips ^[1] — that is, to attach language descriptions to stretches of video after the fact. Training mapped raw pixels and text commands to continuous actions end to end with a standard regression loss ^[1]. Figure said the result was trained with less than 5% of the data volume previously collected for vision-language-action datasets ^[1].

Onboard compute and multi-robot use

Helix runs entirely onboard embedded low-power GPUs, with each robot carrying dual embedded GPUs and System 2 running as an asynchronous background process ^[1]. Running the whole model on the robot rather than on external servers changes what the control stack depends on, since the fast loop does not have to wait on a network round trip.

Figure also said a single set of Helix weights can run simultaneously on two robots collaborating on a shared long-horizon task ^[1]. Because the weights are shared and require no per-task tuning ^[1], the same model can be loaded onto more than one body instead of being retrained for each machine or each job.

Where it sits

Helix belongs to a line of work that treats robot control as a sequence problem — pixels and words in, actions out — rather than as a stack of separately engineered modules. Its distinguishing choices, as Figure described them, are the separation of a slow vision-language model from a fast visuomotor policy ^[1], the breadth of the 35-degree-of-freedom action space ^[1], the very different sizes of the two networks ^[1], and the decision to run everything on onboard hardware ^[1]. Those choices also define what has to be checked in use: a single weight set with no task-specific fine-tuning ^[1] is only useful if it generalises across tasks, and a two-rate control loop is only useful if the fast half really can act on what the slow half passes it.