What is sim-to-real transfer?

Why simulate at all

Because reality is slow, expensive and breakable. A robot learning to walk by reinforcement learning needs to fall over hundreds of thousands of times, and no hardware budget survives that. In simulation it costs nothing, runs faster than real time, and repeats exactly.

The unlock was massively parallel simulation: thousands of copies of the robot stepping simultaneously on one GPU, turning years of walking experience into an afternoon. Nearly every impressive legged-locomotion result of the past few years rests on it, including Figure's learned natural walking, which replaced a hand-tuned controller with an end-to-end network.

What the simulator gets wrong

The sim-to-real gap is the accumulated set of small lies. Friction is approximated. Contact — the exact moment a foot or finger meets a surface — is the hardest thing to model and the thing robotics cares about most. Motors have delay. Gears have backlash. Cables stretch. Sensors are noisy in ways nobody bothered to characterise. The floor is slightly more slippery today than yesterday.

None of these matters much alone. Together they mean a policy that exploited a quirk of the simulator — and policies are extremely good at finding quirks — meets a world where the quirk is absent.

The two standard fixes

They pull in opposite directions and are usually used together.

Domain randomization varies the simulation deliberately: friction, masses, latency, lighting, sensor noise, all shuffled every episode so the policy cannot depend on any of them being exact. The aim is that reality arrives looking like one more variation it already handles. The cost is conservatism — a policy robust to everything is rarely optimal for anything.

System identification goes the other way: measure the real robot's actual parameters and feed them back so the simulator matches this machine. Narrower and sharper, but it must be redone per robot, and it drifts as hardware wears.

One makes the gap not matter; the other makes it smaller.

How far it has closed

Far enough to be visible. Boston Dynamics' electric Atlas carrying a 100-pound fridge was reported as evidence of a closing sim-to-real gap and a shift toward robust whole-body physical AI — that class of contact-rich, whole-body loading is exactly what used to survive only in simulation. Researchers have pushed into stylistic subtlety too: AdaPT transfers the tennis styles of Federer, Nadal and Djokovic onto humanoids.

The tooling has become an industry of its own. Genesis AI pitches Genesis World 1.0 as turning the sim-to-real gap into a compute problem, moving evaluation out of slow physical labs. Unitree adopted Reborn's open-source Roboverse infrastructure. At the cheap end, Hugging Face released a 2,500-dollar 3D-printed humanoid explicitly aimed at closing the gap for people without lab budgets — and has since been acquired by NVIDIA in a 12.9-billion-dollar deal that pulls the field's open-source plumbing into the largest simulation vendor's orbit.

Where it still does not work

Locomotion transfers well because the physics that matter — gravity, ground reaction, momentum — are the physics simulators model best. Manipulation transfers badly, because it is contact all the way down: deformable objects, friction at the fingertip, tactile feedback that no engine reproduces convincingly.

That asymmetry explains a lot of the field's shape. Walking is largely learned in simulation; grasping is still largely learned from teleoperation on real hardware. Any claim that the gap is closed should be read as a claim about one of those, not both.