Our approach
Psychologists describe two modes of thought: a fast, automatic System 1 and a slow, deliberate System 2. Robots need both — and they need to know which one is in charge.
S1 · Classical robotics
Classical robotics is fast, predictable and verifiable. State estimators, controllers and planners run at high rate, on modest embedded hardware, with behaviour you can reason about and test. That is the layer that keeps a vessel on course and a rover on its path — and it is the layer we never hand over to a model whose output cannot be checked.
S2 · Frontier models
Foundation models bring what classical robotics lacks: semantic understanding of a scene, and the ability to break a long-horizon mission into steps. Vision-language models read the world; large language models decompose tasks; vision-language-action models suggest what to do next.
In our architecture a learned model’s output is proposed intent, not a command. System 1 is the logical governor: it checks every proposal against physics, sensor evidence and safety limits, refines it, or refuses it. The result is a robot that benefits from foundation models without inheriting their failure modes.
Sense · Think · Act
Sense · S1Geometric state estimation turns raw sensor streams into a trusted state, many times a second.
Think · S1High-frequency, deterministic logic makes the immediate trajectory adjustments a moving robot needs.
Act · S1The logical governor: every command is checked against physics and safety before a motor moves.
Sense · S2Vision-language models add meaning: not just “an obstacle at 40°”, but what it is and why it matters.
Think · S2Large language models decompose long-horizon missions into hierarchical tasks the robot can execute.
Act · S2Vision-language-action models propose intent — a suggestion System 1 is free to refine or refuse.
How the bridge is built
Anything a learned model contributes is weighed against everything else the robot knows, through the same checks as any other source. That is how System 2 can inform System 1 without overruling it.
The layer that moves the machine runs on modest embedded hardware, at high rate, with behaviour you can reason about and test. We keep it that way, whatever sits above it.
We describe what was measured and name what is simulated. An autonomy stack that overstates its certainty is more dangerous than one that admits it is unsure. In that spirit: what we ship today is System 1. The System 2 layer is our roadmap, and the architecture above is built so it can arrive without rewriting what already works.
Contact
We work with builders, integrators and researchers who want classical guarantees and foundation-model capabilities in the same machine.
Emailc.argueta@s1s2.ai