Vision-Language-Action Models

The foundation model for robots

A vision-language-action (VLA) model takes what a robot sees and is told, and outputs physical action — motor commands, trajectories. Where a language model predicts the next word, a VLA predicts the next move. It is the action layer of embodied AI.

Definition

What a VLA model is

A vision-language-action model is an AI foundation model with three ports: vision in (cameras), language in (an instruction or context), and action out (motor commands or robot trajectories). It is the embodied-AI equivalent of a large language model.

The shift from LLM to VLA is the shift from describing the world to acting in it. An LLM that is asked to pick up a cup can write a perfect paragraph about it. A VLA moves the arm.

How it works

Slow reasoning, fast reflexes

Most modern VLA models use a dual-system architecture that mirrors how humans deliberate versus react:

System 2 — slow

A vision-language model understands the scene and the goal. It reasons about what to do — "the cup is on the left, grasp it by the handle." This runs at a few hertz, like deliberate human thought.

System 1 — fast

A diffusion or flow-matching transformer produces the low-level motor commands at 50–200Hz — the closed-loop control that keeps the hand steady and the grip firm. This is the reflex layer.

Real-time motor control at 200Hz cannot wait for a cloud round-trip, which is why VLA inference runs on-device. The hard constraint is data: robot demonstrations are physically expensive to collect, so VLA models lean on internet-scale vision-language pretraining and synthetic demonstrations to fill the gap.

Comparison

VLA vs LLM vs LAM

Three model types, three kinds of output. The clean way to keep them apart: an LLM writes, a LAM clicks, a VLA moves.

DimensionLLMLAMVLA
InputTextText + app/UI stateVision (cameras) + language + proprioception
OutputTextDigital actions — clicks, API calls, form fillsPhysical action — motor torques, trajectories
EnvironmentNone (pure language)Digital — apps and interfacesPhysical world — 3D space, objects, gravity
Origin of termResearchMarketing (Rabbit, 2024)Research (DeepMind RT-2, 2023)
Error costLow — reversibleMedium — mostly reversibleHigh — irreversible physical consequences
Maturity (2026)MatureContested / loosely definedActive research, early deployment
ExamplesGPT, Claude, GeminiRabbit r1 (contested)RT-2, pi0, GR00T N1, Helix, OpenVLA

benned Kin

A VLA decides how. Kin decides what it knows.

A vision-language-action model is the action layer — it turns a goal into motion. But it does not know your home, your people, or how your business does the work. That is a different layer: the knowledge layer.

Kin is benned's personal AI entity — the memory a robot inherits the moment it connects. The VLA supplies the skill; Kin supplies the context. Two layers, one machine: the action model decides how it acts, the knowledge layer decides what it knows.

For businesses, Lore captures what your best people know while they work — so a new machine running a capable VLA still starts at day one, not at zero.

FAQ

Common questions

What is a vision-language-action (VLA) model?

A VLA model takes visual input (cameras) and language input (an instruction) and outputs physical action (motor commands or trajectories). It is the foundation model for robotics — the embodied-AI equivalent of an LLM. Where an LLM predicts the next word, a VLA predicts the next action.

What is the difference between a VLA and an LLM?

An LLM takes text and outputs text, in a purely digital space. A VLA takes images and language and outputs physical action, in the real world, in real time, with irreversible consequences. An LLM writes; a VLA moves.

What is the difference between a VLA and a large action model (LAM)?

A LAM (Rabbit, 2024) acts in digital environments — operating apps and interfaces. A VLA acts in the physical world through a robot body. LAM is largely a marketing term overlapping with AI agents; VLA is the established research term for the action layer of embodied AI. LAM clicks; VLA moves.

How does a VLA model work?

Most use a dual-system architecture: a slow vision-language model reasons about the goal, and a fast diffusion or flow-matching transformer produces motor commands at 50–200Hz. It is trained on scarce robot demonstration data, supplemented by internet-scale pretraining and synthetic demonstrations.

Last updated: July 2026