Vision-Language-Action Model

A Vision-Language-Action (VLA) model is an AI architecture that fuses visual perception, language understanding, and physical action generation into a single system, primarily developed for robotics and autonomous driving. In the robotics space, Astribot and research collaborators proposed the CLAP framework, a two-stage VLA approach that bridges unlabeled human video with annotated robot trajectories through cross-modal alignment, then trains separate models for task planning (CLAP-NTP) and precise control (CLAP-RF) . Star Dust Intelligence's Lumo-1 takes a different path, using an embodied VLM, cross-embodiment joint training, and real-world reasoning-action training on its S1 robot to convert "cognitive" capabilities into full-body manipulation .

On the autonomous driving side, DeepRoute.ai has made VLA its central technical bet. The company began developing a VLA model on Nvidia's Thor chip in 2024 , and by mid-2025 had launched a VLA model jointly with Volcano Engine featuring four core capabilities: "perspective eye," "know-it-all," "translator," and "responsive spirit" . Its self-developed VLA now powers the DeepRoute IO 2.0 ADAS platform, which the firm released in August 2025 . DeepRoute's CEO Zhou Guang has framed VLA as the path toward "Road AGI" .

Not all practitioners accept the VLA orthodoxy. Chen Yilun, founder of an embodied intelligence startup backed by Linear Capital, explicitly rejected the mainstream VLA approach—"first get VLM from LLM, then build VLA on top"—arguing that it reduces robotics to a downstream branch of multimodal AI rather than an independent discipline. His team instead built AWE (AI World Engine), which prioritizes recording physical quantities like time, space, and force over "retina-style" visual expression .