SafeShift: Safety-Informed Distribution Shifts for Robust Trajectory Prediction in Autonomous Driving

As autonomous driving technology matures, safety and robustness of its keycomponents, including trajectory prediction, is vital. Though real-worlddatasets, such as Waymo Open Motion, provide realistic recorded scenarios formodel development, they often lack truly safety-critical situations. Ratherthan utilizing unrealistic simulation or dangerous real-world testing, weinstead propose a framework to characterize such datasets and find hiddensafety-relevant scenarios within. Our approach expands the spectrum ofsafety-relevance, allowing us to study trajectory prediction models under asafety-informed, distribution shift setting. We contribute a generalizedscenario characterization method, a novel scoring scheme to find subtly-avoidedrisky scenarios, and an evaluation of trajectory prediction models in thissetting. We further contribute a remediation strategy, achieving a 10reduction in prediction collision rates. To facilitate future research, werelease our code to the public: github.com/cmubig/SafeShift