XPENG's Innovative Breakthrough with Peking University
XPENG has marked a significant milestone in autonomous driving technology by collaborating with Peking University to develop FastDriveVLA. This novel visual token pruning framework empowers autonomous driving AI to operate with impressive efficiency by focusing solely on essential information. This innovation has resulted in a remarkable 7.5x reduction in computational load, enabling vehicles to ‘drive like a human’.
Recognition at AAAI 2026
FastDriveVLA has gained acclaim after being accepted by AAAI 2026, a prestigious event in the artificial intelligence realm known for its rigorous selection process and a low acceptance rate of just 17.6%. This recognition underlines the quality and potential impact of the research conducted by XPENG and Peking University.
Pushing Boundaries in Autonomous Driving
This achievement reinforces XPENG's position as a leader in AI-driven mobility and accelerates the industry's ability to deploy next-generation autonomous driving solutions efficiently. The development of FastDriveVLA is particularly significant as it highlights the growing trend of using Vision-Language-Action (VLA) models, which have become integral in understanding complex driving scenarios.
The Science Behind FastDriveVLA
These VLA models are adept at encoding visual data into extensive visual tokens, which form the basis of how AI perceives the environment and makes informed driving decisions. However, the proliferation of visual tokens can lead to substantial computational overhead, hampering both speed and real-time responsiveness critical for autonomous vehicles.
FastDriveVLA addresses these challenges by adopting an innovative approach to visual token pruning. It draws inspiration from human drivers who efficiently focus on relevant foreground visuals while disregarding unnecessary background data. The framework incorporates an adversarial foreground-background reconstruction strategy that enhances its ability to sift through visual tokens effectively.
Notably, during tests on the nuScenes autonomous driving benchmark, FastDriveVLA produced outstanding results across various pruning ratios. Significantly, the framework maintained high levels of planning accuracy even when reducing the number of visual tokens from 3,249 to 812, thereby achieving an impressive reduction in processing demands.
XPENG's Continued Commitment to Innovation
This marks the second occasion this year that XPENG has achieved recognition at a leading global AI conference. Earlier in the year, the company stood out as the sole Chinese automaker invited to present at CVPR WAD, showcasing developments in autonomous driving foundation models.
Furthermore, at its AI Day event, XPENG unveiled the VLA 2.0 architecture, which eliminates the traditional ‘language translation’ step, thereby enabling direct Visual-to-Action generation. This breakthrough reimagines the conventional V-L-A pipeline, positioning XPENG at the forefront of AI innovation in mobility.
XPENG's remarkable achievements reflect its robust in-house capabilities throughout model architecture design, training, and deployment. Looking forward, XPENG is steadfast in its mission to reach Level 4 (L4) autonomy, striving to seamlessly integrate advanced AI systems into vehicles. The company's ultimate goal is to offer safe, efficient, and intuitive driving experiences worldwide.
About XPENG
XPENG is dedicated to shaping the future of mobility through continual technology advancements and innovation, identifying itself as an 'Explorer of Future Mobility.' The company is headquartered in Guangzhou, China, and operates R&D centers across multiple cities, including Beijing, Shanghai, Shenzhen, Zhaoqing, and Yangzhou. Additionally, it has established intelligent manufacturing bases in Zhaoqing and Guangzhou.
With a global approach to research, development, and sales, XPENG maintains an R&D center in the United States and multiple subsidiaries in various European nations. The company emphasizes a fully integrated in-house development strategy concerning intelligent driver assistance software and core hardware, thereby providing users with exceptional driving and riding experiences.
XPENG made headlines on August 27, 2020, by listing on the New York Stock Exchange (NYSE: XPEV), achieving significant financial milestones in its IPO for the global new energy vehicle sector. On July 7, 2021, it further expanded its presence by listing on the Hong Kong Stock Exchange (HKEX: 9868), marking a historic dual primary listing for a Chinese new-energy automaker.
Contacts:
For Media Enquiries: Alison Liang, XPENG PR Department
Email: [email protected]
Frequently Asked Questions
What is FastDriveVLA?
FastDriveVLA is an innovative visual token pruning framework developed by XPENG and Peking University, designed to enhance the efficiency of autonomous driving AI.
How has FastDriveVLA been recognized?
It has been accepted for presentation at AAAI 2026, a leading artificial intelligence conference with a highly competitive acceptance rate.
What are the benefits of using FastDriveVLA?
FastDriveVLA significantly reduces computational load, enhancing the real-time performance of autonomous driving systems by allowing AI to focus on relevant visual information.
What advancements did XPENG present at AI Day?
At the AI Day, XPENG introduced the VLA 2.0 architecture, which eliminates the language translation step for a more direct Visual-to-Action generation method.
What is XPENG's goal for the future?
XPENG aims to achieve Level 4 autonomy in driving technology, focusing on safe and efficient integration of AI systems into vehicles for enhanced user experiences.
