Reinforcement learning neural networks grounded in learned visual entities

This invention describes an artificial intelligence system that learns to perform tasks by interacting with an environment. Instead of directly processing raw images, it first identifies distinct visual objects or features within those images. The system then uses these identified visual entities to understand its surroundings and decide what actions to take, making its learning more structured. The claims specify that this visual entity identification uses a neural network to encode image data and then quantizes these encodings into specific embedding vectors.