Explainers

What is a vision-language-action model? How robots turn words into movement

By Humanoid News Staff · Published 9 November 2026 · Last updated 9 November 2026

Short answer: A vision-language-action (VLA) model is an AI model that looks at camera images, reads an instruction such as "put the apple in the bowl", and outputs the motor commands a robot needs to do it. Most VLAs start from a vision-language model trained on web images and text, then learn actions from robot data. Google DeepMind's RT-2, Gemini Robotics, Physical Intelligence's pi0 and Figure's Helix are the best-known published examples.

Where the idea came from. Google DeepMind described RT-2 on 28 July 2023 as a VLA that learns from both web and robotics data. It was built on existing vision-language models (PaLI-X and PaLM-E) and trained with robot demonstrations that DeepMind says were collected with 13 robots over 17 months in an office kitchen. The trick, according to DeepMind, was to represent robot actions as tokens in the model's output, the same way a language model outputs words.

Why it matters. A model that has learned from the web already knows what an apple, a drawer or a recycling bin is. DeepMind says RT-2 could follow commands it had not seen in its robot data and do basic reasoning, such as picking a rock as an improvised hammer. That is the main promise of VLAs: fewer hand-programmed skills and fewer demonstrations per task.

Different designs. Physical Intelligence said on 31 October 2024 that its pi0 model starts from a 3-billion-parameter vision-language model and adds continuous action outputs through flow matching, a variant of diffusion models, so it can send motor commands up to 50 times a second. Figure's Helix, announced on 20 February 2025, splits the job in two: a slower onboard vision-language model running at 7 to 9 Hz for understanding, and a fast visuomotor policy that outputs actions at 200 Hz. Figure says Helix controls the whole humanoid upper body, including individual fingers, and runs on embedded GPUs on the robot. NVIDIA's open GR00T models use a similar "System 1, System 2" split; see our GR00T explainer.

Gemini Robotics. On 12 March 2025 Google DeepMind introduced Gemini Robotics, a VLA built on Gemini 2.0 with physical actions added as an output, and Gemini Robotics-ER for spatial reasoning. DeepMind said it was partnering with Apptronik to build humanoid robots with Gemini 2.0, and that Gemini Robotics more than doubled performance on its generalisation benchmark compared with other state-of-the-art VLAs, by its own measure.

What VLAs still need. Every one of these models is trained partly on robot demonstrations, which are slow and costly to collect. Figure says Helix used about 500 hours of teleoperated data. How companies gather that data, from remote operators to simulation, is covered in how humanoid robots learn.

← All news