The Reality Gap

Policies that achieve 95% success in simulation often drop to 30–50% in the real world. This "reality gap" stems from visual differences (rendering vs real images), physics differences (contact model inaccuracies), and sensor noise patterns. Bridging this gap systematically is one of the most important practical skills in robot learning.

Three Strategies

Domain randomization: train across wide parameter distributions so the real world is "just another sample." System identification: carefully measure and calibrate simulation parameters to match reality. Fine-tuning: train in simulation, then collect a small real-world dataset (50–200 episodes) and fine-tune. Most successful deployments combine all three.

  • Domain randomization: best for visual policies
  • System identification: best for contact-rich tasks
  • Real-world fine-tuning: almost always improves performance
  • Recommended: combine all three

Practical Tips

Start with MuJoCo or Isaac Sim (both have good contact models). Randomize camera position, lighting, and object textures aggressively. Measure real joint friction and damping. Validate sim-to-real on a simple task before attempting complex ones. SVRC provides calibrated simulation models for OpenArm.