Scaling robot learning often requires combining real-world demonstrations with simulation data. Sim-and-real co-training is a common strategy, but it typically treats the simulator as a fixed, unchanging source of data.
The paper "Active Generative Simulation" argues that this fixed treatment is a limitation. Instead of assuming what to simulate, the method meta-learns the simulation strategy itself, deciding which scenarios or data would be most useful for policy training.
By making the simulation process adaptive, the approach aims to improve the efficiency and effectiveness of co-training. The abstract does not provide experimental details, but the core idea is to turn simulation selection into a learned component.