The Sim-to-Real Gap

A policy trained on thousands of GPU-parallel rollouts achieves 100% success in simulation. Deploy it on a real robot — and it falls apart. This sim-to-real gap (reality gap) is the most fundamental challenge in robot learning. Here is where it comes from, how to narrow it, and how to measure it honestly.

Where the Gap Comes From

The differences between simulation and reality accumulate from multiple sources.

Rendering diagram…

Simulation is clean; reality is messy. Friction coefficients are subtly off, camera images carry noise, commands incur latency, motors have backlash. Each deviation is small in isolation, but if the policy has overfit to simulation's cleanliness, these small differences compound into failure.

How to Narrow It — Domain Randomization

The primary remedy is domain randomization: deliberately make the simulation diverse — varying friction, mass, lighting, and noise on every episode — so the policy learns to work "under any conditions." The real world then appears as just another variation, and the gap shrinks.

💡 Put reality inside the distribution — The intuition behind domain randomization is to ensure the real-world parameters fall within the randomization range. If friction is randomized over 0.3–1.2 and the real-world friction is 0.7, the policy has seen that value and is not surprised. If the range fails to cover reality, the gap remains.

Measure It Honestly

The most important part — measuring the gap honestly.

⚠️ Do not present "simulation success" as "validation complete" — In our actual records, one sim2real manipulation configuration passed all simulation KPIs while achieving 0% real-world grasp success. Reporting it as "working" from simulation numbers alone would have been a lie. In sim-to-real work, truth is the real-world success rate — not the simulation success rate.

Expressed numerically: sim2real_gap = (sim KPI) − (real-world KPI). Tracking this gap openly reveals what to randomize further and what needs better modeling. Hiding it — looking only at simulation numbers — means never closing it.

In a nutshell

📌 The sim-to-real gap accumulates from physics, sensor, latency, and actuator mismatches. Domain randomization narrows it by putting reality inside the training distribution — but truth is the real-world success rate. Passing simulation KPIs is not "validation complete" (as one case with a 0% real grasp rate demonstrated). Track the gap as a number; never hide it.