FPD-Drive: Future Policy Distillation for End-to-End Autonomous Driving

Overview

FPD-Drive transfers privileged future knowledge into a visual driving policy to improve both trajectory generation and selection. A teacher uses recorded future observations during training to provide diverse, high-quality trajectory targets. A student learns from these targets and from evaluated safety outcomes of its own candidates. At deployment, the student plans using four current camera views, available ego-state history, and a navigation command, without future observations or teacher inference.

A logged driving scene contains only one executed trajectory and its associated future observations, leaving other valid behaviors weakly supervised. FPD-Drive addresses this limitation with two complementary objectives:

• Future-Privileged Multi-Target Distillation expands supervision beyond a single demonstration by transferring multiple safe, geometrically distinct teacher trajectories.
• Safety-Aware Score Correction improves candidate selection by correcting overconfident unsafe predictions and increasing confidence in underestimated safe alternatives.

All scores are on a 0–100 scale; higher is better. FPD-Drive-Pro denotes the scaled model.