Training one robot network instead of thousands through factored learning

Robotic control tasks involve multiple independent factors—object identity, obstacles, colors—but collecting demonstrations for every combination explodes combinatorially. This work trains a single shared diffusion network with dropout per factor, so its score function decomposes additively across factors at inference. On drone racing through gates, the approach passes 90% of held-out combinations (matching an oracle) versus 3% for multi-network baselines, and transfers zero-shot to new venues with 2.4× fewer crashes.