For a cost-prohibitive example, you could put a humanoid robot whose pretraining contained no chemistry, in a chem lab, and put it in a hypothesis loop, where it generates a random action, guesses the result, and then does it. Every time the lab burns down, you just build it again. Reward the model for producing novel actions and reward the model for correct hypotheses.

In other words, you generate more training data by actually experimenting, just like humans do.