Ralf Römer

We propose FIPER, a framework to predict failures of generative imitation learning policies by detecting consecutive OOD observations and high action uncertainty, achieving more accurate and faster detection than prior approaches.