←Home

    When Language Models Take the Wheel

    How vision-language-action models turn pretrained multimodal language models into end-to-end driving planners that perceive, reason, and output trajectories, and how they are trained, sped up, made safe, and tested.

    Background: Vision-language-action models

    In order

    1 stories
    Boston Dynamics Spot 外形可辨, 無標誌
    Boston Dynamics Spot 外形可辨, 無標誌

    Photo: Mika Baumeister

    Embodied · Driving VLA models

    Qwen-Drive adds 3D perception and a planner to a 4B VLM, keeping its architecture

    Qwen researchers build on a 4-billion-parameter vision-language model to map scenes in 3D, answer driving questions and plan paths, with mixed results across benchmarks.

    arXiv preprint · 31 August 2026