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

What a VLA actually does

Three things arrive at the model and one thing leaves. In go camera frames — usually several views, wrist and head — and a natural-language instruction: put the mug in the dishwasher. Out come actions: joint angles or end-effector targets, produced continuously as the scene changes.

What makes that notable is what it replaces. The conventional robotics stack broke the same job into stages — perceive the scene, identify the mug, plan a grasp, plan a path, execute — each hand-built, each with its own failure modes, and the whole chain re-engineered for every new task. A VLA collapses the stages into one learned function. Nobody writes the grasp planner.

The name is literal, and the order matters: it is built by taking a vision-language model — the architecture behind image-understanding chatbots — and extending it with a third capability, producing physical motion. That inheritance is the whole point.

Why it moved robotics onto foundation-model timelines

A robot built the traditional way knows what its engineers taught it. A VLA starts from a model that has already absorbed an enormous amount of the internet, so it arrives knowing what a dishwasher is, that mugs have handles, and that tidy up implies things belong somewhere. None of that had to be programmed.

That is why the field's clock changed. Robotics had spent decades on task-specific systems; VLAs let it inherit progress from a field moving far faster, and the pace of announcements since has followed language-model rhythms rather than mechanical-engineering ones.

The catch is the part that cannot be inherited. Text is scraped; actions are not. Every VLA still needs demonstrations of physical behaviour, which is why so much of the industry's effort goes into teleoperation rigs, motion capture and egocentric video — and why sample efficiency is the field's central economic question.

Who is shipping one

Google DeepMind has been the most public. Gemini Robotics 1.5 reframed the robot as a "physical agent" that reasons and plans rather than following commands, and Gemini Robotics 2 pushed the same line from tabletop manipulation into full-body humanoid control and multi-robot teams. A separate on-device version runs locally on the robot, trading model size for the latency and reliability you get from not depending on a network.

Physical Intelligence is the closest thing the field has to a pure play — a San Francisco lab building vision-language-action foundation models for hardware other people make. Its RL Tokens work had robots mastering screwdriving and zip-tying in about fifteen minutes, eventually beating human speed, which goes straight at the standard criticism of VLAs: fluent at coarse motion, hopeless at the last millimetre.

Figure has taken the opposite route, keeping model and robot in one company. Helix 02 is a single network that lets the Figure 03 walk and manipulate at the same time, collapsing the walk-then-act sequencing that made earlier humanoid demos look so stilted.

Two more worth knowing. Sharpa's MoDE-VLA uses a mixture-of-experts model to reach contact-rich bimanual tasks, demonstrated by peeling an apple. Skild AI's S1 performs unseen ten-minute tasks from a single video prompt with no task-specific training.

Watch the acronyms, though. Xpeng badges its IRON humanoid's system as a "VLT" — vision-language-task. Vendors rename the category routinely, and the rename rarely marks a different architecture.

The case against

The most interesting thing about VLAs in 2026 is that the frontier has started defining itself in opposition to them.

The objection is about data. Producing physical behaviour by imitating demonstrations means needing demonstrations of everything, and Yann LeCun has argued that the sheer data requirements behind current humanoid demos are evidence the approach will not scale, predicting a paradigm shift toward world models — systems that learn to predict consequences rather than copy actions.

That is no longer only a position in interviews. NVIDIA's DreamZero is explicitly framed as going beyond the VLA, using video diffusion to give a robot something like physical imagination. Dyna Robotics makes the claim numerically, arguing that a world-action model trained on a million hours of human video establishes a human-to-robot scaling law that challenges VLA dominance.

Whether that is a successor or a component of the same stack is genuinely unsettled.

How to read a VLA claim

Four questions worth asking of any announcement: