What a robot needs is a prediction of the world's next moment

With the release of a technique that rebuilds a 3D space from just a few photos taken from different positions and places a virtual camera inside it, a conclusion takes shape: world models are not a special-effects technique but a precondition for physical AI. If a language model predicts the next word, a world model predicts how the world will change in the next moment, and for a robot one more thing is added: how the world will change if I act this way. The limit is contact: weight, friction and hardness cannot be known through a camera, and while action decisions are made tens of times per second, contact requires hundreds to more than 1,000 hertz, so today's robots effectively ignore contact and handle only objects whose properties can be anticipated. It also draws a distinction: grasping the same cup 10,000 times to raise the success rate is optimization for specific conditions, while grasping 10,000 objects never seen before is an entirely different problem of generalization. The proposed solution is a loop: capture the site, reconstruct the space, train a virtual robot thousands of times while varying the lighting and positions, then transfer it to the real machine.

This is a note: my own write-up after reading several sources. It has no source link; events checked against a source are kept as records.