MVVLP

Autonomous Valet Parking (AVP) represents a critical application of autonomous driving; however, existing approaches remain constrained by limited scene understanding, insufficient instruction comprehension, and the absence of standardized multimodal benchmarks. To address these limitations, this work introduces the Multi-View Vision-Language Parking (MVVLP) dataset, a dedicated benchmark for Vision-and-Language Navigation (VLN) in parking scenarios. MVVLP incorporates multi-view image acquisition, fine-grained semantic annotations, and diverse natural language instructions reflecting realistic user expressions. The dataset spans four indoor parking facilities and five distinct instruction styles, thereby enriching the input space for VLN tasks and enhancing multimodal semantic understanding and decision-making. In addition, we develop the VLP Simulator, which enables autonomous agents to interact with realistic parking environments by simulating vehicle position transitions. Experimental evaluations confirm that MVVLP provides an effective and discriminative platform for assessing VLN methods within AVP contexts, addressing critical gaps in realism, diversity, and reproducibility. Collectively, MVVLP and the simulator establish a comprehensive foundation for advancing robust, scalable, and generalizable decision-making systems in autonomous parking applications.