XPENG VLA2.0: A Foundation Model for Physical AI

Part of the Special ECE Seminar Series

Modern Artificial Intelligence

Title:

XPENG VLA2.0: A Foundation Model for Physical AI

Speaker:

Xianming Liu, University of Illinois Urbana-Champaign

Abstract:

This talk will introduce Xpeng VLA 2.0, the industry's first mass-produced, billion-parameter-scale foundation model for physical AI. We will explore its novel "Vision-Implicit Token-Action" architecture, an end-to-end approach that bypasses language bottlenecks to directly translate visual data into action. Trained on large scale unlabeled, real-world video clips, VLA 2.0 functions as a self-evolving generative action and world model, demonstrating emergent abilities to understand complex gestures and navigate untrained scenarios. We will also cover the full-stack optimization for its on-vehicle deployment and its function as a unified platform for Xpeng's entire ecosystem, from autonomous vehicles to robotics. The system will be massively deployed to all customer vehicles in Q1 2026.

Bio: