Imagine learning to swim. If someone throws you into the deep end on day one, you'll panic and struggle. But if you stay in the shallow end forever, you'll never learn to handle waves or currents. Training robots for grasping faces the exact same dilemma. Domain randomization the standard approach for bridging the sim-to-real gap randomly varies parameters like object position, friction, or action noise during training. The idea is simple: expose the robot to enough variation, and it will generalize to the real world. But here's the flaw: most implementations use a fixed randomization level from start to finish. Early in training, when the robot hasn't even learned to reach properly, high variability creates noisy, unstable reward signals. Later, when the robot has improved, the same fixed level may no longer push it toward better generalization. In our experiments, a fixed randomization baseline trained for 500,000 steps never reached a 70% success rate. It plateaued at around 50%.
Adaptive Curriculum Domain Randomization for Robotic Grasping: A Performance-Driven Approach
Most robotic grasping policies fail in the real world because simulation training uses fixed difficulty. Too hard early on destabilizes learning; too easy later limits generalization. We propose ACDR—difficulty adjusts based on success rate. Result: 74.5% avg success vs. 33.2% (+41.3 pp).