机器人公司实测 Astra 做视觉到动作控制
Frame Robotics 称 Astra 让机器人控制重新像软件问题:看摄像头、失手后下一轮自动改抓取角度,但单次循环约 15 秒。
GPT-6 Astra is the first model that makes robot control look like a software problem again. We've been trying for years to make robots usable outside the lab. last couple of months that work actually moved. not a cheaper arm. a model that looks at the camera, misses, and changes grip angle and carry height on the next try. Nvidia cosmos folks say it cracked RoboLab near-perfect. people who lived in VLAs say they have not seen this kind of image-to-action understanding before. the old line was "LLMs cannot control robots." the new line is "which setup." It is still slow. ~15 second loops, the arm sitting idle while the model thinks, then flying blind until the next cycle. hardware still matters. a 48-task robotics-engineering bench exists because control is not the whole job. But training and eval got cheap enough that more teams can actually run them. next months you will see more of that land in boring use cases, not just lab clips. that is the part that was supposed to take a decade. We are confident our ecosystem will play a huge role in that.