Why simulation carries most of the training load
Collecting real robot demonstration data is slow and expensive. A human teleoperating a physical robot arm can generate maybe a few hundred demonstrations in a working day, and every hour of real robot time requires real hardware, physical setup, and someone supervising for safety. Simulation removes those bottlenecks. A modern physics simulator can run thousands of parallel simulated robots, each attempting a task with slightly randomized objects, lighting, and physics parameters, generating years of equivalent experience in hours on a cluster of GPUs. This is what makes training data volumes comparable to what deep learning needs at all feasible for robotics.
This technique is often called domain randomization: by exposing the model to a wide spread of simulated variation in textures, object shapes, friction coefficients, and camera angles, the training process forces the model to learn features that are robust to nuisance variation rather than overfitting to the exact pixel statistics of one simulated environment. The hope is that if a policy performs well across a wide enough distribution of simulated worlds, it will also perform reasonably in the one real world it eventually gets deployed into.