VLA models explained

What is a vision-language-action model?

A vision-language-action (VLA) model is a neural network that takes in images and a natural-language instruction, and directly outputs robot commands: joint positions, velocities or torques. Where classical robotics chained separate modules (object detection, trajectory planning, control), a VLA learns the entire pipeline end to end, from demonstrations.

The term took hold in 2023 with Google DeepMind's RT-2, which showed that a vision-language model pretrained on web data transfers its world knowledge to controlling a robot arm: the robot inherits object recognition, common sense and instruction understanding for free. Since then, the VLA has become the foundation of generalist manipulation, and the focus of intense competition between labs.

Where do VLA models come from? They are the outcome of a longer history: our history of AI in robotics pillar traces how earlier waves, programmed automation and then learned perception, laid the groundwork that today's generalist models build on.

How it works

A VLA is made of two blocks. The backbone is a vision-language model (VLM) pretrained on huge web corpora: it provides semantic perception and instruction understanding. The action head converts the backbone's representations into motor commands.

Two approaches dominate for this head. The first discretises actions into tokens, treated as extra vocabulary of the language model: the choice of RT-2 and OpenVLA, simple but limited in precision and control frequency. The second generates continuous actions through diffusion or flow matching: Physical Intelligence's π0 and NVIDIA's GR00T produce smooth bursts of actions (action chunks) at high frequency. GR00T N1.6, for instance, relies on a 32-layer diffusion transformer for this module.

Training follows the large-model recipe: pretraining on heterogeneous data (multi-robot demonstrations, human video, synthetic data), then fine-tuning on the target robot and tasks. A notable 2026 development: π*0.6 adds a reinforcement learning stage (the RECAP method) that lets the model learn from its own failures in production.

How does a VLA differ from an LLM?

A VLA shares the DNA of large language models: a transformer architecture and massive vision-language pre-training. Most VLAs actually embed an LLM or VLM as the "brain" that understands the scene and the instruction. But everything else diverges, because the physical world is not a document.

This difference in kind explains the dominant 2026 architecture: a slow reasoning module (System 2, often a VLM) paired with a fast action head that executes (System 1). The LLM brings world understanding; the VLA adds what no text contains: motion.

The 2026 landscape

The landscape has taken shape in eighteen months: every major AI player now has its own VLA, and Europe entered the race with Mistral AI in July 2026.

Two takeaways from this table. First, specialisation has begun: Robostral Navigate targets navigation while Helix and GR00T aim at generalist manipulation. Second, the size spread is striking: from 450 million parameters (SmolVLA) to several billion, a sign that the right model size is still an open question. Find these models and their specs in our models comparator.

Dual-system architectures

A single network cannot simultaneously reason about a complex instruction and drive motors at 200 Hz. Hence dual-system architectures, inspired by the distinction popularised by Daniel Kahneman: a slow System 2, a VLM that understands the scene and plans at a few hertz, and a fast System 1, a reactive policy producing motor actions at high frequency.

Figure's Helix popularised this scheme: a VLM running at 7-9 Hz guides a visuomotor policy at 200 Hz. Helix 02, unveiled in early 2026, adds a System 0, a whole-body balance and coordination layer running at kilohertz rates. Google DeepMind splits the roles across two distinct models: Gemini Robotics-ER reasons, plans and calls digital tools, while Gemini Robotics 1.5 executes. GR00T N1.6 follows the same logic with Cosmos Reason as the slow brain and its diffusion transformer as the fast controller. This convergence towards a slow/fast hierarchy is one of the clearest architectural consensuses in the field.

Current limits

Three limits shape the debate in 2026.

• Generalisation. Most VLAs are evaluated in environments close to their training data. π0.5 was designed precisely for open-world generalisation (new kitchens, new homes), but transfer to genuinely novel environments remains the hard part.
• Control frequency. The bigger the backbone, the slower the inference. Action chunks and dual-system architectures work around the problem, at the cost of reduced reactivity to disturbances between plans.
• Evaluation. There is no standard benchmark: each lab publishes its own protocols, real-robot evaluations are expensive, and cross-model comparisons remain fragile. Initiatives such as RoboChallenge are trying to fix this with large-scale evaluations on real robots.

Add to this the question of reliability: a 90% success rate is fine for a demo, not for a production line. That is the gap that real-world reinforcement learning methods such as RECAP aim to close.

Open or closed weights: the industrial stakes

For an industrial company, the choice between open weights and closed models is not ideological, it is operational. Open weights (GR00T, OpenVLA, SmolVLA, π0 via openpi) can be fine-tuned on your cells, your parts and your data, run on premises without shop-floor images ever leaving the factory, and can be audited. A closed model is consumed through an API or a partnership, with the dependency that implies.

The players' strategies are easy to read. NVIDIA opens its models to become the base layer of the ecosystem, what TechCrunch sums up as "the Android of generalist robotics": the model is open, while the compute that trains and runs it remains its core revenue. Figure keeps Helix proprietary and vertically integrated with its robots. Google DeepMind opens access to reasoning (ER via API) but reserves the VLA for selected partners.

For European industry, the question doubles as a sovereignty issue: Mistral AI's entry into robotics opens a continental alternative. The pragmatic advice: prototype on open weights to keep control of your data and your roadmap, and maintain an abstraction layer so you never depend on a single vendor. Compare the robots that embed these models in our robots comparator.

Keep reading

• Place the VLA within the wider stack with our physical AI stack pillar.
• Understand the other model family that complements the VLA: our world models and sim2real guide.
• Compare VLA models in production in our models comparator.