Tuesday, December 30, 2025

Xpeng and Peking College paper accepted by AAAI

Xpeng and Peking College have developed FastDriveVLA, a framework that reduces computational load in autonomous driving programs

Xpeng and Peking College have developed FastDriveVLA, a visible token pruning framework for autonomous driving that has been accepted for presentation at AAAI 2026, one of many main synthetic intelligence (AI) conferences globally with an acceptance charge of 17.6% this 12 months. The framework reduces computational load by roughly 7.5 occasions whereas sustaining planning accuracy, addressing a key problem in deploying vision-language-action (VLA) fashions for real-time end-to-end autonomous driving programs.

VLA fashions encode photographs into visible tokens that allow autonomous programs to interpret their environment and make driving selections. Nonetheless, processing massive numbers of those tokens will increase computational calls for, affecting inference pace and real-time efficiency in autos.

FastDriveVLA makes use of a reconstruction-based strategy impressed by how human drivers deal with related foreground info akin to lanes, autos, and pedestrians whereas filtering out non-critical background areas. On the nuScenes benchmark, the framework decreased visible tokens from 3,249 to 812 whereas sustaining excessive planning accuracy.

This marks Xpeng’s second recognition at a serious AI convention in 2025, following a presentation at CVPR WAD in June on autonomous driving basis fashions. In November, the automaker unveiled its VLA 2.0 structure, which removes the language processing step to allow direct visual-to-action technology.

Supply: Xpeng

Related Articles

LEAVE A REPLY

Please enter your comment!
Please enter your name here

Latest Articles