Its not all that different, we take multiple asynchronous streams of messages integrate them into a coherent predictive model, and use that to feed the reinforcement learning. The messages can contain images, a simple case can be seen in the demo with a 1d vision sensor.