simulation-to-reality transfer (sim-to-real)

Training a robot directly in the real world is slow, expensive, and breaks hardware. So we train millions of times faster inside a physics simulator and then move the learned policy onto the real machine — much like a pilot who logs hours in a flight simulator before ever flying a plane.

Simulation is cheap, massively parallel, safe, and hands you perfect state labels for free. The catch is the reality gap: simulators get friction, contact, latency, and sensor noise subtly wrong, so a policy that looks flawless in sim can fail on hardware. The standard remedies are domain randomization, system identification, domain adaptation, and a little fine-tuning on real data.

Sim-to-real works best when the mismatch lives in parameters you can randomize or measure. It struggles when the simulator is missing whole phenomena — deformable cloth, complex fluid flow, unusual contacts — that the policy therefore never trained against.