VLA: The Brain, the Eyes, the Hands

The next trillion-dollar AI company probably won't build a chatbot. It will build a model that can see your kitchen, hear "clean up," and physically do it. That model already exists, it worked in real apartments last year, and the architecture behind it is surprisingly simple to understand.

I'll be honest, when I first stumbled on the term "VLA" a few months ago, I thought it was yet another acronym the AI hype machine invented to sound smart. Vision-Language-Action models. Sounds like something someone puts on a slide deck to justify a $600 million funding round.

Then I actually dug into the research and something clicked, because VLAs represent the clearest progression I've seen in AI right now, and almost nobody outside the robotics community is paying attention.

Here's what rewired my thinking.

Your LLM is a brain in a jar. It can reason, plan, and compose, but it has no eyes, no hands, and no body. It lives entirely in the world of text. Researchers fixed this in two steps: first they gave it eyes (vision encoders that project images into the same token space the LLM already understands), and then they gave it hands (action outputs that generate motor commands from what it sees and hears).

That's the LLM → VLM → VLA progression in one paragraph. A brain, then a brain that sees, then a brain that sees and acts in the physical world.

Now here's where it gets interesting: the field can't agree on how to represent physical actions. One camp (Google's RT-2, OpenVLA) discretizes robot movements into tokens, same vocabulary trick as language. The other camp (Physical Intelligence's π₀) argues actions should stay continuous and uses flow matching to generate smooth trajectories directly. Both work, and the most interesting research right now is happening at their intersection.

What made me take this seriously wasn't the tech demos though, it was the investment signals. Physical Intelligence raised $600M in their Series B, VLA papers at ICLR went from 9 to 164 in one year (that's 18×), and π₀.5 was tested in completely unseen San Francisco homes last April, not labs, actual apartments it had never seen, and it worked.

The pattern underneath VLAs (extend a foundation model to new modalities and new output types) is the same pattern powering AI agents today. If you're building systems that take actions in digital environments, you're solving a version of the same architectural problem.

Resources if you want to go hands-on:

→ OpenVLA (open-source 7B, beat Google's 55B RT-2-X) → π₀ (open-sourced by Physical Intelligence) → NVIDIA GR00T N1 (humanoid VLA foundation model) → Open X-Embodiment (1M+ real robot trajectories, 34 labs) → HuggingFace LeRobot (the framework to start with) → UnifoLM-VLA — Unitree, open-source, built for the G1 → VLA 2.0 — XPeng, shared across robots, cars and drones on the same model