CUHK's Li Hongsheng Team Paper MindVLA-U1: VLA No Longer Loses to VA, Language Truly Enters Autonomous Driving Decision-Making
Researchers from the Chinese University of Hong Kong, Li Hongsheng's team, have developed MindVLA-U1, a unified architecture for autonomous driving that integrates visual, language, and action (VLA) components. This new model aims to overcome the limitations of previous VLA approaches, which often struggled with planning accuracy and real-time performance, by enabling language understanding to directly influence driving decisions. MindVLA-U1 achieves this through an architecture that processes continuous video streams with memory, uses language-predicted driving intents to guide trajectory generation, and can switch between fast and slow reasoning paths for efficiency and complex scenario handling. AI
IMPACT Enables autonomous driving systems to move beyond reactive visual processing to proactive decision-making based on semantic understanding.