Federated Reinforcement and Imitation Learning for Safe and Scalable Autonomous Driving
Title: Federated Reinforcement and Imitation Learning for Safe and Scalable Autonomous Driving
DNr: Berzelius-2026-361
Project Type: LiU Berzelius
Principal Investigator: André Teixeira <andre.teixeira@it.uu.se>
Affiliation: Uppsala universitet
Duration: 2026-11-17 – 2027-06-01
Classification: 10210
Keywords:

Abstract

This project is carried out by Adam Rokah, an industrial PhD student at Zenseact affiliated with Uppsala University through WASP. It develops privacy-preserving, safety-aware planning models for autonomous driving that are trained across a fleet of vehicles without centralizing raw sensor data. Supervisors are André Teixeira and Per Mattsson (Uppsala University), and Mina Alibeigi and Maciej Wozniak (Zenseact). The approach combines Federated Imitation Learning (FIL) and Federated Reinforcement Learning (FRL). Simulated clients first learn from human driving data through imitation learning, and the policies are then refined with reinforcement learning for long-horizon performance. Clients share only model updates. Later stages of the project add robustness to non-IID client data, safety assurance through constrained RL and runtime safety filters, and data quality control inspired by Byzantine-robust aggregation. The first project year covers three parts: a position paper on federated fleet learning for autonomous driving (a taxonomy of centralized, hybrid and federated architectures, first-order scaling laws, and a roadmap towards fleet-scale federated learning); a literature review with baseline implementations; and a first federated imitation learning pipeline for planning, starting from object-level perception outputs and extending, where feasible, to end-to-end planners such as VAD and SSR, evaluated in simulation. Later years add RL refinement, safety mechanisms and larger-scale experiments. Methods are evaluated in several driving simulators (e.g., CARLA) and on real-world driving datasets and benchmarks such as the Zenseact Open Dataset (ZOD), NAVSIM and nuPlan.