Meta Muse 自己装 PyTorch 让小车找到黑球
有人让 Meta Muse 控制一台消费级小车找黑球，它自己装了 PyTorch 和 Depth Anything V2 做深度估计并开过去，全程无人写这条流水线。
An AI agent was told to find a black ball. It only had a 2D camera, so it installed PyTorch and Depth Anything V2 inside its own VM, estimated the distance, sanity checked its own number, then drove the rover there. Nobody wrote that pipeline for it. Meta Muse assembled its own perception-to-action stack on an Earth Rover Mini+, and Alexandr Wang posted about it watching from the sidelines. Raj ran the experiment on @BitRobotNetwork hardware, which is the detail that makes it more than a demo. The Mini+ is a cheap consumer rover sitting on a sidewalk, not a lab platform behind a reservation system. An agent capable of writing its own perception layer paired with hardware anyone can buy is a very different distribution curve than agentic robotics on a $50,000 research arm. The RoboCap stack got sharper the same week. Grounded Superintelligence came out of stealth with an API doing metric 3D hand tracking and sub centimeter depth at manipulation range, integrated into LeRobot and Rerun, built specifically for the RoboCap suite. Raw egocentric video into policy ready data in minutes. One experiment by one person is not a trend, and agentic reliability across messy environments is still unmeasured. But the moment a machine installs its own tools to satisfy an instruction, the line between operating a robot and asking one stops being a metaphor.