i make most of my post graphics with ai image tools. they can draw any fountain or street you ask for, because people have been posting photos of those online for 20 years. ask the same models to understand the inside of a random warehouse well enough for a robot to walk through it, and there's barely anything to learn from. fei-fei li has been saying this for a while. robotics doesn't have the internet that language models had, and data is the bottleneck. her answer starts with capturing the real world, then building variety in simulation from there. funny thing is she's been here before. imagenet happened because thousands of people labeled photos by hand. @vangrid_io feels like the 3d version of that idea. people get paid to go film the exact place a model is missing, and every capture gets fingerprinted and anchored on base. crowds built the dataset that taught ai to see. maybe they build the one that teaches it to move around too. what's a place near you that no map has ever really seen?