AI controller learns to drive any robot using only one camera and no sensors. After 2–3 h of random wiggling recorded by 12 RGB-D cameras, the system maps visible 3-D points to actuators via a “visuomotor Jacobian field,” letting the robot see itself, predict motion and self-correct like a human learning finger control. Cheap, sensor-free and occlusion-robust, it outperforms 2-D camera baselines and works for novel architectures without retraining.