One of the recent focuses of AI research is End-to-End Autonomous Driving, as it is gaining interest due to its promising potential. However, it has an unexplored limitation: driving performance decreases when an E2E driving AI model is deployed in ve...
One of the recent focuses of AI research is End-to-End Autonomous Driving, as it is gaining interest due to its promising potential. However, it has an unexplored limitation: driving performance decreases when an E2E driving AI model is deployed in vehicles that differ from those used in training, which we denote as vehicle domain gap. To address this problem, this paper introduces the framework MVAdapt, a novel foundation model for effective adaptation to various vehicle archetypes. The main contribution is a low-level multi-head transformer encoder that generates physics-integrated scene feature embeddings. By deriving physical properties of the target vehicle, MVAdapt dynamically decodes target vehicle-specific waypoints with real-time performance. Even in an unseen vehicle that has extreme physical differences, MVAdapt proves its few-shot adaptation ability, as it quickly closes the performance gap by leveraging a tiny dataset and fine-tuning with it. The evaluation experiments performed with the CARLA Leaderboard 1.0 benchmark demonstrate that MVAdapt outperforms state-of-the-art performance across various metrics. Our work presents a practical method for achieving robust cross-vehicle generalization, pioneering the way for more scalable and adaptable autonomous driving AI. Moreover, we propose a new research area for autonomous driving, vehicle-domain adaptation.