I think it’s a combination of simulation, YouTube videos, and specially recorded training footage. The last one is expensive, but given the funding these startups receive, I’m pretty sure they can scale their RL methods at least 10x.