视觉语言动作模型Vision-Language-Action Model
Vision-Language-Action Model
让一个模型同时看图、读指令、出动作，端到端控制机器人。Let one model see images, read instructions and output actions to control a robot end to end.
视觉语言动作模型（VLA）把视觉编码器、语言模型与动作输出接在一起：输入相机画面和自然语言指令，直接输出机器人该执行的动作。
RT-2 把网络图文知识迁移到机器人控制，显示这类模型能把没见过的指令泛化成新动作；OpenVLA 给出了开源的同类实现。
它与「感知—规划—控制」分模块搭机器人的传统路线相对：一个模型从看到动，中间不再手写每个环节。
A vision-language-action model (VLA) connects a visual encoder, a language model and action output: camera images and natural-language instructions go in, and robot actions come out directly.
RT-2 transfers web image-text knowledge into robot control and shows such models can generalise unseen instructions into new actions; OpenVLA provides an open-source implementation of the same kind.
It contrasts with the traditional modular perception-planning-control route: one model goes from seeing to acting without hand-writing every intermediate stage.
这是当前具身智能里最受关注的模型形态，新闻里的「机器人大模型」多指它。知道它端到端的本质，就能追问两个边界问题：训练数据覆盖不到的任务怎么办，动作出错时谁来兜底。This is the most watched model form in embodied AI; robot large model in news often means it. Knowing its end-to-end nature lets you ask two boundary questions: what happens on tasks outside training coverage, and who catches action errors.
VLA · VLA 模型
回到名词地图看第2站的其他 43 个词 →Back to the map: the other 43 terms in station 2 →