Personalized Federated Learning of Driver Prediction Models for\n Autonomous Driving

Abstract

Autonomous vehicles (AVs) must interact with a diverse set of human drivers\nin heterogeneous geographic areas. Ideally, fleets of AVs should share\ntrajectory data to continually re-train and improve trajectory forecasting\nmodels from collective experience using cloud-based distributed learning. At\nthe same time, these robots should ideally avoid uploading raw driver\ninteraction data in order to protect proprietary policies (when sharing\ninsights with other companies) or protect driver privacy from insurance\ncompanies. Federated learning (FL) is a popular mechanism to learn models in\ncloud servers from diverse users without divulging private local data. However,\nFL is often not robust -- it learns sub-optimal models when user data comes\nfrom highly heterogeneous distributions, which is a key hallmark of human-robot\ninteractions. In this paper, we present a novel variant of personalized FL to\nspecialize robust robot learning models to diverse user distributions. Our\nalgorithm outperforms standard FL benchmarks by up to 2x in real user studies\nthat we conducted where human-operated vehicles must gracefully merge lanes\nwith simulated AVs in the standard CARLA and CARLO AV simulators.\n