Helix has a fast brain and a slow brain. So does my golf swing
Figure's Helix splits robot control into a slow model that understands and a fast one that moves. A golf swing works the same way, which is why thinking during it ruins it.
Figure announced Helix last week, a vision-language-action model that controls the whole upper body of its humanoid robot: arms, wrists, individual fingers, head and torso. In the demo, two robots are handed groceries they’ve never seen, and they put them away in a kitchen, passing items to each other and figuring out where things go from a spoken instruction.
The architecture is what caught my attention. Helix has two systems.
System 2 is a 7-billion-parameter vision-language model, pretrained on internet data, that runs at 7 to 9 Hz. It looks at the scene and the instruction and figures out what to do: this is a ketchup bottle, it goes in the fridge door, the other robot is holding out a bag. It outputs a latent vector that encodes the plan.
System 1 is an 80-million-parameter visuomotor policy that runs at 200 Hz. It takes the camera images, the robot’s joint states and System 2’s latent vector, and outputs continuous commands for 35 degrees of freedom. It does the actual moving, fast enough to adjust grip and posture in real time.
The two are trained end to end together, and the naming is from Kahneman’s Thinking, Fast and Slow: fast, automatic control and slow, deliberate reasoning. Previous vision-language-action models either ran a big model slowly, which made motion jerky and slow, or ran a small fast policy that didn’t understand much. Splitting them lets each run at the speed its job needs.
I’ve been taking golf lessons seriously over the last while, and this is exactly the lesson I keep learning the hard way.
A golf swing takes about a second from takeaway to impact. The downswing is around a quarter of a second. Good players have a consistent tempo, often roughly 3 to 1 between backswing and downswing, and I’m working toward something closer to that from my rushed 2.5. There is no time in the downswing to think “rotate the hips, keep the lead wrist flat, don’t cast.” If you try, you get stiff and late, and the ball goes somewhere unfortunate. The swing has to be run by something fast and automatic that was trained by thousands of repetitions.
The slow system does its work before the swing: reading the lie, choosing a club, picking a target, setting up. Then it hands over a compact intention, “smooth 8-iron, slight draw, at the flag,” and gets out of the way. My coach’s advice to have one swing thought at most is really advice about bandwidth between the two systems. The slow system is allowed to send one small latent vector. More than that and it interferes with the fast one.
I suspect robotics and golf arrive at this split for the same reason. Any system that has to act physically in real time, with a body that has dynamics, has to separate slow understanding from fast control, and train the fast part until it doesn’t need supervision. Figure’s robot learned that from data. I’m learning it one bucket of balls at a time.