- Breakthrough in XPENG-PKU research: XPENG, in collaboration with Peking University, developed FastDriveVLA, a new visual token A pruning framework that allows self-driving AI to “drive like a human” by focusing only on essential information, achieving a 7.5x reduction in computational load.
- High-level AI recognition: The research was accepted by AAAI 2026, one of the world’s leading AI conferences, which had a very selective acceptance rate of only
17.6% This year. - Accelerate L4 autonomy: This achievement highlights XPENG’s comprehensive capabilities in AI-driven mobility and advances the industry toward efficient and scalable deployment of next-generation autonomous driving systems.
The document presents FastDriveVLA, an effective visual tool token Pruning framework specially designed for Vision-Language-Action (VLA) end-to-end autonomous driving models. This work proposes a new approach to the visual token pruning by allowing AI to “drive like a human,” focusing only on essential visual information while filtering out irrelevant data.
As large AI models evolve rapidly, VLA models are widely adopted in end-to-end autonomous driving systems due to their strong capabilities in complex scene understanding and action reasoning. These models encode images into a large number of visual tokens, which serve as the model’s basis for “seeing” the world and making driving decisions. However, processing a large number of tokens increases the computational load onboard the vehicle, impacting inference speed and real-time performance.
Although visual token Pruning has been recognized as a viable method for accelerating VLA inference, existing approaches whether based on textual-visual attention or token similarity, showed limitations in driving scenarios. To solve this problem, XPENG and PKU developed FastDriveVLA, a new reconstruction-based system. token Pruning framework inspired by the way human drivers focus on relevant foreground information while ignoring non-critical background areas.
The method introduces an adversarial foreground and background reconstruction strategy that improves the model’s ability to identify and retain valuable tokens. On the nuScenes autonomous driving benchmark, FastDriveVLA achieved peak performance across different pruning ratios. When the number of visual tokens was reduced from 3,249 to 812, the framework achieved a nearly 7.5x reduction in computational load while maintaining high scheduling accuracy.
This is the second time this year that XPENG has been recognized at a leading global AI conference. In June, XPENG was the only Chinese automaker invited to speak at CVPR WAD, where it shared its advances in core autonomous driving models. At its AI Day in November, XPENG unveiled the VLA 2.0 architecture, which removes the “language translation” step and enables direct generation from visual to action, a breakthrough that redefines the conventional VLA pipeline.
These achievements reflect XPENG’s comprehensive in-house capabilities, from model architecture design and training to vehicle distillation and deployment. Looking ahead, XPENG remains committed to achieving L4-level autonomous driving to accelerate the integration of physical AI systems into vehicles, aiming to bring safe, efficient and comfortable intelligent driving experiences to users around the world.
About XPENG
XPENG is committed to leading the transformation of future mobility through technological exploration, positioning itself as “Explorer of Future Mobility”. Based in
XPENG pursues a global research, development and sales strategy, with an R&D center in
On August 27, 2020, XPENG was officially listed on the New York Stock Exchange (NYSE: XPEV), raising funds in an IPO that set a record at the time for the global new energy vehicle industry. On July 7, 2021, the company was listed on the Hong Kong Stock Exchange (HKEX: 9868), becoming the first Chinese new energy automobile manufacturer to achieve a dual primary listing in both countries.
For more information, please visit https://www.xpeng.com/.
Contacts:
For media inquiries: Alison Liang, XPENG Public Relations Department
E-mail: liangrq3@xiaopeng.com
Show original content:https://www.prnewswire.com/news-releases/xpeng-peking-university-collaborative-research-accepted-by-aaai-2026-introducing-a-novel-visual-token-pruning-framework-for-autonomous-driving-302650038.html
SOURCE XPENG