All terms

Glossary · AI & ML

Sim-to-real transfer

Also known as: sim2real

In brief

Sim-to-real transfer is the technique of training a policy in simulation, then deploying it on a physical robot with minimal additional fine-tuning. Domain randomization — varying the simulator's physics, visuals, and sensor noise during training — is the standard tool for closing the sim-to-real gap.

Simulation is orders of magnitude cheaper than real-world rollouts: free episodes, no hardware wear, no human safety risk, parallelizable across thousands of GPUs. The catch is that policies trained in a perfect simulator often fail on real hardware because the simulator doesn't capture every physical detail. Latencies, friction, sensor noise, and actuator non-linearities all diverge from the sim.

The standard fix is domain randomization: train in many simulators with different parameters so the policy learns to be robust to the variation. The hope (often borne out) is that the real world is "just another sample" from the distribution. Zero-shot sim-to-real is the gold standard — deploy the policy on hardware with no real-world fine-tuning. Figure's S0 perception policy on Figure 03 is a recent public example.

Related terms

See it in the wild

Browse robots and brands using these techniques

Glossary entries are upstream. The catalog is where the implementations live.