MVAdapt: Zero-Shot Multi-Vehicle Adaptation for End-to-End Autonomous Driving
Summary: arXiv:2604.11854v1 Announce Type: cross
Abstract: End-to-End (E2E) autonomous driving models are usually trained and evaluated with a fixed ego-vehicle, even though their driving policy is implicitly tied to vehicle dynamics. When such a model is deployed on a vehicle with different size, mass, or drivetrain characteristics, its performance can degrade substantially; we refer to this problem as the vehicle-domain gap.
To address it, we propose MVAdapt, a physics-conditioned adaptation framework for multi-vehicle E2E driving. MVAdapt combines a frozen TransFuser++ scene encoder with a lightweight physics encoder and a cross-attention module that conditions scene features on vehicle properties before waypoint decoding.
Key Features of MVAdapt
- Integration of TransFuser++: MVAdapt utilizes a frozen TransFuser++ scene encoder to retain the high-level understanding of the driving environment.
- Lightweight Physics Encoder: This component ensures that vehicle dynamics are incorporated into the adaptation process, allowing the model to adjust effectively to different vehicle configurations.
- Cross-Attention Module: By conditioning scene features on vehicle properties, this module enhances the model’s ability to decode waypoints accurately, even in diverse scenarios.
Performance Evaluation
In the CARLA Leaderboard 1.0 benchmark, MVAdapt demonstrates substantial improvements over two baseline approaches: naive transfer and multi-embodiment adaptation. The evaluation shows that MVAdapt excels on both in-distribution and previously unseen vehicles.
Complementary Behaviors
MVAdapt exhibits two critical complementary behaviors:
- Strong Zero-Shot Transfer: The model showcases impressive capabilities in adapting to a variety of unseen vehicles without needing additional training.
- Data-Efficient Few-Shot Calibration: For vehicles that represent severe physical outliers, MVAdapt enables effective calibration using minimal data, ensuring reliable performance across a range of conditions.
Conclusion
The results from the MVAdapt framework suggest that explicitly conditioning end-to-end driving policies on vehicle physics is a significant advancement towards developing more transferable autonomous driving models. This research opens new avenues for enhancing the robustness and versatility of autonomous driving systems across diverse vehicle types.
For those interested in exploring the technical details further, all associated codes are available at
MVAdapt GitHub Repository.