Sim-to-real transfer

Sim-to-real transfer is the practice of training a robot control policy in simulation and then deploying it on physical hardware, with the goal of having the simulated learning carry over to real-world performance without significant additional training. The strict version of the claim is zero-shot transfer: the policy that runs on the real robot is identical to the one trained in simulation, with no real-world fine-tuning. Looser versions involve simulation-pretraining followed by real-world supervised fine-tuning, which is a meaningfully easier task.

The distinction matters because the technique is increasingly invoked as a milestone, but the strict and loose versions reflect very different engineering claims. A press release that mentions sim-to-real without specifying whether real-world fine-tuning was used should be read as the loose version by default.