Survey maps four years of vision-language-action models for robots
An arXiv survey reviews 88 works on models that turn web-pretrained vision systems into robot control, tracing design, training and deployment trends.
In plain words
Vision-language-action (VLA) models adapt vision-language models pretrained on web data so they can output robot actions, letting web knowledge carry over to physical control. The survey follows these models through their life cycle, from design to deployment, and identifies trends in training data, model size and generalist capability.
What was announced
The survey, covering December 2022 to October 2026, reviews 88 works focused mainly on manipulation and navigation, plus humanoid control, world-action models and evaluation. It analyses 22 models across seven axes spanning model design, training and deployment, reporting training data, model size, inference speed and hardware. The authors report three trends: training data has broadened from roughly 10,000 hours of teleoperated robot data to community-collected episodes and egocentric human video; model sizes have split, with most recent models at 3-7 billion parameters while a new group under one billion runs on consumer GPUs; and generalist VLAs handling both manipulation and navigation are emerging.
The authors also compare VLAs with world-action models, which predict future observations alongside actions. They review evidence that some VLAs with near-perfect benchmark scores fail when objects or the camera move slightly, and they open further research questions. The field has moved quickly from RT-2, the first VLA, to open models fine-tunable on a single consumer GPU and proprietary generalists that clean unseen homes.
What it does not show
- The source is the paper's abstract only, so the survey's full findings, methods and evidence cannot be verified from this excerpt.
- Benchmark fragility is reported as reviewed evidence, not independently confirmed by the survey or this news item.
Sources
Source-support review completed 17 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 17 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.