Dual-System Architecture (System 1 / System 2)

A slow, deliberate large model handles understanding and planning, while a fast, lightweight model handles real-time motor control.

The dual-system architecture borrows psychologist Daniel Kahneman's language from Thinking, Fast and Slow: System 2 is slow and deliberate, System 1 is fast and intuitive. Applied to robots, System 2 is usually a vision-language model with a few billion to a few tens of billions of parameters, running a few to a dozen times per second, looking at the scene and understanding the instruction, and outputting a semantic goal or a latent vector; System 1 is a much smaller policy network outputting joint actions at up to hundreds of hertz. This combines a large model's common sense with the low-latency demands of real-time control. Representative systems are Figure's Helix (February 2025) and NVIDIA's GR00T N1 (March 2025). Figure's Helix 02, from January 2026, adds a 1kHz System 0 for balance and contact, becoming a three-system architecture.

ExampleFigure Helix's S2 is a 7-billion-parameter open-source VLM running at 7 to 9Hz; its S1 is an 80-million-parameter Transformer controlling 35 degrees of freedom across the wrists, fingers, torso, and head at 200Hz. GR00T N1's vision-language part runs at 10Hz on an L40 GPU, and its action part outputs actions at 120Hz.

Also called

System 1 / System 2, S1 / S2, Fast-Slow System, Fast System, Slow System, System 0/1/2 Architecture

Related

Embodied Reasoning Model · Hierarchical Architecture · Vision-Language-Action Model · Figure Helix · NVIDIA Isaac GR00T N1 · System 0

Sources

Helix: A Vision-Language-Action Model for Generalist Humanoid Control (Figure)
Helix 02 (Figure)
GR00T N1: An Open Foundation Model for Generalist Humanoid Robots (arXiv 2503.14734)

As of

2026-01