Sim-to-Real: Why Robots Trained in Simulation Still Struggle in the Real World

Training a robot in a physics simulator is fast and cheap. Getting that training to survive contact with reality is the hard part.

Training a humanoid's control policy directly on real hardware is slow, expensive, and hard on the robot — every fall is a repair bill. So most locomotion and manipulation policies are trained largely or entirely in simulation first, then transferred to the real machine. That transfer step is where a lot of otherwise-promising research quietly stalls.

What actually breaks

Contact dynamics are the hardest thing to simulate accurately: friction, restitution, and how deformable or compliant surfaces behave under load. A policy that looks flawless in simulation can fail immediately on real ground because the simulator's contact model didn't match reality closely enough.

Sensor noise, actuator delay, and manufacturing variance between individual robots of the same model all add further mismatch a policy never saw during training.

The tools of the trade

Domain randomization — deliberately varying friction, mass, sensor noise, even visual textures during training — is one of the more reliable ways to make a policy robust to the mismatch between sim and reality, though overdoing it slows convergence and can produce overly conservative behavior.

NVIDIA's Isaac Lab is widely used for large-scale parallel training thanks to GPU-accelerated physics, while MuJoCo is often used as a secondary validation step, since its contact dynamics are generally considered more realistic even though it trains more slowly at scale. A common workflow trains in one and validates in the other before ever touching hardware.

"Zero-shot" rarely means zero

You'll see the phrase "zero-shot sim-to-real transfer" a lot. In practice it rarely means literally no real-world tuning — it usually means the policy transfers well enough to be usable with calibration and safety-limit adjustments, rather than requiring a whole second training phase on hardware. That's still a real achievement; it's just a smaller claim than the phrase suggests.