A DVDrive Approach for doScenes Instructed Driving Challenge

The paper presents a submission to the doScenes Instructed Driving Challenge, adapting the OmniDrive model, a vision-language-action driving agent, to predict ego trajectories based on instruction-annotated nuScenes scenes. Key innovations include a DVPE-style divided-view perception module that enhances multi-view visual grounding by performing visibility-aware cross-attention within localized view spaces, reducing irrelevant cross-view interference. This advancement is significant for practitioners as it improves the alignment of natural language instructions with relevant visual data in autonomous driving applications.