MVAdapt: Zero-Shot Multi-Vehicle Adaptation for End-to-End Autonomous Driving

Researchers Haesung Oh and Jaeheung Park have introduced MVAdapt, a novel framework that addresses a critical limitation in current autonomous driving systems: the 'vehicle-domain gap.' Traditional end-to-end (E2E) driving models are trained on data from a specific vehicle with fixed dynamics (size, mass, drivetrain). When deployed on a different vehicle, performance degrades significantly. MVAdapt solves this by explicitly conditioning the driving policy on vehicle physics, allowing a single model to adapt to various vehicles without full retraining.

The system architecture cleverly combines a frozen, pre-trained TransFuser++ model—which handles scene understanding from sensor data—with a new, lightweight physics encoder. A cross-attention module then conditions the scene features on the specific vehicle's physical properties before generating driving waypoints. This approach enables both zero-shot transfer to many unseen vehicles and data-efficient few-shot calibration for extreme physical outliers, making deployment more flexible and cost-effective.

In rigorous testing on the CARLA Leaderboard 1.0 simulation benchmark, MVAdapt demonstrated superior performance compared to naive transfer methods and other multi-embodiment adaptation baselines. The results validate that explicitly incorporating vehicle dynamics into the AI's decision-making process is a crucial step toward more generalizable and robust autonomous driving systems. The team has made all code publicly available, accelerating further research in this direction.