Particle.news
Download on the App Store

XPENGPeking University’s FastDriveVLA Accepted at AAAI 2026, Cutting VLA Compute Nearly 7.5x

The reconstruction-based token pruning method reports state-of-the-art nuScenes results with far fewer visual tokens.

Overview

  • The peer-reviewed paper was selected for AAAI 2026, which accepted 4,167 of 23,680 submissions for a 17.6% rate.
  • FastDriveVLA reduces visual tokens from 3,249 to 812, delivering a nearly 7.5-fold cut in computational load while maintaining planning accuracy.
  • The approach uses adversarial foreground–background reconstruction to retain critical cues such as lanes, vehicles, and pedestrians.
  • Reported performance is on the nuScenes benchmark and is presented as enabling more feasible real-time inference for end-to-end VLA driving models.
  • The acceptance follows XPENG’s CVPR WAD appearance and its VLA 2.0 reveal, as the company cites progress toward future Level 4 autonomy.