What is VLA in AI?

What is VLA in AI? | Generated by AI

Home 2026.07

Question: What is the full name of VLA in AI?

Answer: VLA = Vision-Language-Action model.

It is a class of AI models that combine:

Vision (V) → understand images/video from cameras
Language (L) → understand instructions, reasoning, and concepts through natural language
Action (A) → output actions to control an agent, usually a robot

The basic pipeline:

Camera image
     |
     v
[ Vision Encoder ]
     |
     v
[ Multimodal Transformer / LLM ]
     |
User instruction:
"Pick up the red cup"
     |
     v
Action tokens
     |
     v
Robot controller
     |
     v
Motor commands

Compared with LLMs:

Model	Input	Output
LLM	Text	Text
VLM (Vision-Language Model)	Image + Text	Text
VLA (Vision-Language-Action)	Image + Text	Robot actions

Examples of VLA models:

Google DeepMind RT-2 (Robotics Transformer 2)
- Uses vision-language knowledge from large models and converts it into robot actions.
Physical Intelligence π0 (pi-zero)
- A general-purpose robot foundation model.
NVIDIA GR00T (Project GR00T)
- Humanoid robot foundation model.
OpenAI robotics research has also explored multimodal embodied intelligence.

The key idea:

A VLA is basically an LLM for robots: instead of generating the next text token, it generates the next physical action.

Technically, many modern VLAs are built by extending Transformer architectures:

image patches
      |
Vision Transformer (ViT)
      |
      +------ text tokens
      |
      v
Large Transformer
      |
      v
action tokens
      |
      v
trajectory / motor commands

A common training objective is:

[ P(a_t | I, L, a_{<t}) ]

where:

(I) = visual observation
(L) = language instruction
(a_t) = next action

Similar to GPT:

[ P(x_t | x_{<t}) ]

but the predicted token is an action token instead of a word.

References:

Back Donate