VLA-Corrector：40M轻量模块为机器人动作闭环纠错
浙大与达摩院用40M参数模块解决VLA动作分块导致的滞后，以极小代价实现机器人实时纠错，兼顾效率与稳健性。
VLA机器人因动作分块（一次性生成多条动作序列）产生开环盲区，即使相机已捕捉到目标移动，仍会执行旧计划。浙大与阿里达摩院提出VLA-Corrector，在推理旁挂载一个约40M参数模块，实时监控视觉特征偏差，一旦动作过时就中断并触发重规划。
正文摘录
VLA-Corrector: A 40M Corrector Lets Robots Correct Errors While Acting A robot grasps a block and is about to place it into a bowl. Suddenly, someone moves the bowl away. A human would instinctively pause, look around, and re-aim. But many VLA robots don't. Even as the camera feed continues streaming, the robot keeps executing motion commands generated hundreds of milliseconds earlier—eventually dropping the block into thin air. This isn't because the model "didn't see" the change. It's because during that window, the robot was never given the chance to re-plan. VLA-Corrector, a joint effort by the OmniAI lab at Zhejiang University's ACES lab and Alibaba DAMO Academy, targets this blind spot that has long been masked by action chunking. While predicting a sequence of actions in one pass is…