Reproduce “Safety Evaluation of Motion Plans Using Trajectory Predictors as Forward Reachable Set Estimators” on OpenResearch

About this paper

Kaustav Chakraborty, Zeyuan Feng, Sushant Veer, Apoorva Sharma, Wenhao Ding, Sever Topan, Boris Ivanovic, Marco Pavone and 1 more

A runtime safety monitor interprets multimodal trajectory predictions as probabilistic forward reachable sets for checking autonomous-vehicle plans. By combining convex set extraction, conformal calibration, and Bayesian confidence adaptation, the method reduces the conservatism of worst-case reachability while improving robustness to distribution shift; on nuScenes, its hybrid variant reached a 6.77% balanced error rate in the out-of-distribution city setting with 95.07% coverage.

Key insights

• Multimodal predictions can represent alternative agent behaviors more efficiently than a single confidence region: increasing the number of GMM modes from one to five reduced the required covariance inflation and generally lowered balanced error.
• Conformal calibration addresses systematic prediction error under the calibration distribution, but it does not by itself guarantee performance after a distribution shift; online confidence adaptation and worst-case fallbacks improve OOD behavior at the cost of more false alarms.
• The framework uses learned behavior models as data-driven reachability estimates rather than treating them as definitive forecasts, providing a middle ground between overconfident prediction regions and highly conservative formal reachable sets.

Summary by alphaXiv.