Vision-language-action model (VLA)
Also known as: VLA, VLA model, vision-language-action
A vision-language-action model (VLA) is an AI model that combines what a robot sees (vision), instructions in plain language, and the movements needed to act (actions) in one system.
You might tell a robot 'put the apple in the bowl,' and the VLA turns camera input and that sentence into arm and gripper commands. While VLAs have made robots far better at handling new objects and instructions, they still fail in unpredictable ways and need huge amounts of robot training data.
Vision-language-action model (VLA), explained at three levels
At home
Think of a new kitchen helper who has read every cookbook online and watched thousands of people cook, then learns your kitchen by practising. A VLA works like that for a robot. It starts with an AI trained on internet images and text, then learns from recordings of robots doing tasks. Given a camera view and an instruction like 'put the apple in the bowl,' it outputs arm movements.
At work
A vision-language-action model is a single AI model that takes camera images and a plain-language instruction and outputs robot motions. Because it inherits general knowledge from models trained on web data, it can handle objects and requests it was never explicitly programmed for. For businesses, it promises robots that can be retasked through instructions and demonstrations instead of custom code.
At home
A VLA builds on a vision-language model, the kind of AI that can describe a photo. Researchers fine-tune it on recordings of robots performing tasks, pairing each camera frame and instruction with the motion that followed. The model learns to predict the next small movement, repeating many times a second until the job is done. Because it keeps the original model's knowledge, it can follow 'pick up the snack' even for a brand it has never seen.
At work
Most VLAs pair a pretrained vision-language backbone with an action output. Some, like Google DeepMind's RT-2 and OpenVLA, turn movements into tokens, like words. Others, like Physical Intelligence's pi-zero, add an action expert that outputs smooth continuous motions. Newer systems, such as Figure's Helix and NVIDIA's GR00T N1, split into a slower reasoning model and a fast controller running many times a second. Training data comes from pooled datasets such as Open X-Embodiment.
At home
A VLA learns patterns, not rules, so it can generalize to new objects but also fail where a traditional program never would, such as grabbing the wrong item when the lighting changes. It also can't explain its choices, which makes testing hard. While the leap in flexibility is real, in my opinion these models need far more testing than a demo video suggests before they work near people.
At work
The main limits are data, speed, and safety. Robot data is scarce compared with internet text, so teams collect it through teleoperation, simulation, and human video. Large models run slowly, so hierarchical designs hand fast control to smaller models. Policies trained on one robot body transfer poorly to another. There is no accepted way to certify a learned policy, so safety comes from external guards such as speed limits, force sensing, and fenced cells under ISO 10218. I'd measure failure rates on the real task first.
Real-world examples
At home
Aisha visited a university robotics lab where a researcher typed 'hand me the blue cup,' and a robot arm picked the right cup from a cluttered table, despite never seeing that cup before.
At work
Thomas runs automation at a Quebec food packaging plant. He's testing a VLA-powered robot arm that can be taught new packing patterns from a few demonstrations instead of reprogramming.
How it compares
| feature | Vision-language-action model (VLA) | Traditional robot program |
|---|---|---|
| Input | Images and language | Coded instructions and positions |
| New objects | Can generalize | Must be programmed |
| Predictability | Lower | High |
| Setup for new tasks | Examples or instructions | Engineering work |
VLAs trade predictability for flexibility.
For analysts
VLA-based robots can cut reprogramming costs for varied tasks. I'd keep them in fenced or supervised settings until their failure rate on your actual task is measured.
History & origin
Earlier robots relied on hand-written programs for each task. Researchers began training robots on large datasets in the 2010s. Google's RT-1 in 2022 trained a transformer on robot demonstrations, and Google DeepMind's RT-2, announced in July 2023, coined the term vision-language-action model by building on a vision-language model trained on web data. The Open X-Embodiment dataset, released in 2023 by dozens of labs, pooled data from many robot types. OpenVLA followed in 2024 as an open model, Physical Intelligence released its pi-zero (π0) model in late 2024, and in 2025 Figure showed Helix, NVIDIA released GR00T N1 for humanoids, and Google DeepMind launched Gemini Robotics. In my view, VLAs are the first robotics approach that makes the 'general-purpose robot' idea seem plausible, though still years from routine.