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

Key points

• End-to-end autonomous driving models are typically trained on real-world driving logs, which record only the trajectory actually executed in each scene.1
• Specifically, we introduce Future-Privileged Multi-Target Distillation to transfer diverse, high-quality trajectory targets from a teacher with access to recorded future observations to the student.1
• For candidate selection, we further introduce Safety-Aware Score Correction to correct ranking errors between nearby safe and unsafe candidates and raise underestimated safety scores of safe candidates.1
• Our base model achieves 94.9 PDMS on NAVSIM v1 navtest and 55.9 EPDMS on NAVSIM v2 navhard.1

Sources (1)

• [1]FPD-Drive: Future Policy Distillation for End-to-End Autonomous Driving
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 9, 08:50 AM

  We propose FPD-Drive, an end-to-end planner that distills privileged future knowledge into a visual driving policy to improve both multi-trajectory supervision and safety selection.

  End-to-end autonomous driving models are typically trained on real-world driving logs, which record only the trajectory actually executed in each scene.

Extractive summary: sentences quoted from the sources.