Domain Randomization Study

Key Insight

Domain randomization is the primary technique for bridging the sim-to-real transfer gap, training policies in a simulator whose physical parameters are dynamically varied to prevent overfitting to simulated physics. By measuring the policy's success rates in a held-out test simulation with extreme physics values, this study quantifies the performance drop caused by the reality gap and how randomization cures it. The key is that by exposing the policy to a wide distribution of masses, frictions, and latencies during training, the robot learns a robust control strategy that generalizes to physical hardware without requiring system identification.