Dual-System VLAs: Fast and Slow Control

A large vision-language model understands scenes and instructions well, but it can only run a few times per second. Fast reactions, like adjusting a grip as an object slips, need a hundred or more updates per second. Dual-system designs, such as Figure's Helix and NVIDIA's GR00T N1, pair a slow "System 2" that understands with a fast "System 1" that reacts.

Architecture Slow model rate Fast controller rate Target speed

Two rates

The slow system, a vision-language model, runs at $f_{\text{slow}}$, a few hertz. It decides what to do and outputs a goal or a latent command. The fast system, a small policy, runs at $f_{\text{fast}}$, up to hundreds of hertz, using fresh sensor readings each time to follow that command. With only the slow model, the robot acts on information up to $1/f_{\text{slow}}$ old:

$$e_{\text{lag}} \approx v_{\text{target}} \cdot \frac{1}{f_{\text{slow}}} \qquad \text{versus} \qquad v_{\text{target}} \cdot \frac{1}{f_{\text{fast}}}$$

What to notice

With the large model alone, the hand jumps toward where the target was at each slow update, so it trails a moving target. In the dual design, the slow model only says "follow the red ball", and the fast controller tracks the ball itself at high rate, so the error drops sharply. Speed up the target and the gap between the two grows. This is the same split humans use: deliberate thinking sets goals, fast reflexes handle the details.