What is a vision-language-action (VLA) model, and who is building them?

A vision-language-action (VLA) model is a single neural network that takes camera images and a plain-language instruction and returns the motor commands a robot should execute next. It is made by taking a vision-language model, the kind of AI that can describe a photo, and training it further on robot demonstrations so that its output is movement rather than text.

The term entered robotics in July 2023, when Google DeepMind described RT-2 in an arXiv preprint and reported that a model trained on web images and robot data roughly doubled success on unseen objects and scenes against its earlier RT-1 controller, at about 62% versus 32%. Three years later the label covers models from Google DeepMind, Physical Intelligence, NVIDIA, Figure AI, Skild AI and AgiBot, and the companies behind them have raised some of the largest private rounds in robotics. This explainer sets out how the models work, how they differ from older robot software and from world models, who leads, and why the cost of data, more than model design, now decides the commercial race.

How a VLA turns pixels and words into motor commands

Most VLAs have three parts. A vision encoder turns each camera frame into a sequence of visual tokens. A language backbone, the pretrained vision-language model, reads those tokens together with the instruction and the robot's own joint readings. An action head then converts the backbone's internal state into numbers for joints, grippers or wheels. NVIDIA's model card for GR00T N1 lays this out plainly: a SigLIP 2 vision transformer reading 224 by 224 pixel frames, a T5 text encoder, a small network for the robot's state that changes with each robot body, and a diffusion transformer that produces the actions.

Starting from a vision-language model buys transfer. A backbone trained on vast numbers of captioned web images already knows what most household and factory objects look like, so a robot can act on things it never saw during robot training. The RT-2 authors reported that this inherited knowledge let a robot pick the smallest object on a table or move a can onto a number it had only met on the web, with success on such emergent tasks of 60% against 17% for RT-1. The harder question is how the network should express a movement at all.

Tokens or a flow: two ways to output an action

The first VLAs treated movement as words. RT-2 split each of eight action dimensions, six for the gripper's position and rotation, one for opening the gripper and one for ending the task, into 256 bins, so a motion became a short string of tokens that the language model emitted like text, according to the paper. OpenVLA, a 7-billion-parameter model from a Stanford-led academic team released with open weights in June 2024, used the same approach and reported a success rate 16.5 percentage points higher than Google's 55-billion-parameter RT-2-X across 29 tasks.

Tokens are slow and coarse for dexterous work. The largest RT-2 ran at 1 to 3 Hz, its paper says,. Physical Intelligence's π0, described in an October 31, 2024 preprint, attached a separate 300-million-parameter action expert to a 3-billion-parameter PaliGemma backbone and trained it with flow matching, a relative of the diffusion method behind image generators. The expert starts from random noise and refines it over 10 steps into a chunk of 50 future actions, which lets π0 control robots at up to 50 Hz. GR00T N1 does the same job with a diffusion transformer.

A third design separates thinking from reflexes. Figure AI's Helix, announced on February 20, 2025, pairs a 7-billion-parameter vision-language model running at 7 to 9 Hz with an 80-million-parameter policy running at 200 Hz, steering 35 degrees of freedom on two onboard GPUs, the company said. NVIDIA calls the same split System 2 and System 1, and AgiBot's GO-2, released on September 8, 2026, uses an asynchronous dual system in which a slower semantic planner hands action sequences to a faster module that corrects for noise. That layering is also what separates a VLA from the robot software factories already run.

Why a VLA is neither robot programming nor a world model

Classical industrial robots are programmed point by point. An integrator teaches waypoints with a pendant or writes the path offline, and the program repeats unchanged until a part shifts or the line is retooled. A VLA replaces that script with a learned policy that reads the scene every cycle, so it can cope with clutter and new objects, but it gives up the strict repeatability that makes scripted cells easy to certify. Japanese makers are trying to keep both. Yaskawa Electric said in July 2026 that its MOTOMAN NEXT system uses Google DeepMind's Gemini Robotics ER 1.6 to plan tasks from high-level instructions without detailed teaching, while the robot's own vision, path planning and force sensing stay underneath.

A world model answers a different question. It predicts how a scene will change, usually as video, given what the robot does, and it is used to generate synthetic training data or to test policies before they touch hardware. Runway said on September 30, 2026 that policy results inside its world model correlate with real-world results at 0.95. The two ideas are converging. Runway calls its new Praxis-1 a world action model that outputs robot control, and Generalist AI pretrains GEN-1 on more than 500,000 hours of human interaction data recorded with wearables, with no robot data at all, according to the company.

Who builds VLAs, and who gives them away

Google DeepMind launched Gemini Robotics, built on Gemini 2.0, on March 12, 2025 with Apptronik as its humanoid partner, and on July 30, 2026 released Gemini Robotics 2 for whole-body humanoid control; its action models remain limited to partners and trusted testers. Physical Intelligence went the other way. Its openpi repository carries π0, π0-FAST and, since September 2025, π0.5 under an Apache 2.0 licence. NVIDIA published GR00T N1 on March 18, 2025 under a non-commercial licence, and its 3-billion-parameter GR00T N1.7 ships under the NVIDIA Open Model License inside Hugging Face's LeRobot library. Figure keeps Helix in-house, Skild AI sells its Skild Brain and S1 models through deployments, and AgiBot delivers GO-2 through its Genie Studio platform.

Openness is a business decision. Open weights let universities and small hardware makers adapt a model on one workstation; openpi says low-rank fine-tuning needs more than 22.5 GB of GPU memory, within reach of an RTX 4090. Closed models keep the data flywheel and the customer relationship with the owner. Where the two camps really part is in how they pay for data.

Data, not architecture, is the cost line

Unlike chatbots, VLAs cannot scrape their training data from the web. Robot demonstrations are mostly gathered by teleoperation, with a person steering a robot through a task. Figure said Helix learned from about 500 hours of teleoperated behaviour. Physical Intelligence used more than 10,000 hours across seven robot configurations and 68 tasks for π0. Skild AI said in August 2026 that collecting 380 demonstrations of a task longer than four minutes takes 50 to 100 hours of teleoperation, and that its in-context S1 model gets comparable value from a single video.

Companies are attacking that cost from three sides. NVIDIA generated 780,000 synthetic trajectories, equal by its count to about 6,500 hours of human demonstrations, in 11 hours, and said mixing them with real data lifted GR00T N1's performance by 40%. Generalist and General Intuition are betting on human and gameplay video. Figure built Index, a crowdsourced video programme that had gathered 16 million videos by August 2026, and committed $3.5 billion to Nscale for GPUs to train Helix on it.

The cost of moving a model to a new robot is falling as well. Google DeepMind said Gemini Robotics 2 adapts to a new two-armed robot in a few hours, typically with fewer than 200 examples. That is the figure a factory buyer should watch, because it decides whether a model can move from one cell to the next without a fresh data campaign, and it is what investors have been paying for.

Where the money has gone

Skild AI raised $1.4 billion in January 2026 at a valuation of more than $14 billion, Figure AI raised $1 billion at $39 billion post-money in September 2025, and Generalist AI raised $400 million in June 2026, according to the companies. Skild is the only one of the group to put revenue on record, saying on September 10, 2026 that it had passed $100 million in annual recurring revenue with more than 60 paying customers.