VECTOR-Drive: Tightly Coupled Vision-Language and Trajectory Expert Routing for End-to-End Autonomous Driving

VECTOR-Drive introduces a tightly coupled vision-language-action framework for autonomous driving that balances semantic reasoning with motion planning through expert routing. Built on Qwen2.5-VL-3B, the system achieves 88.91 Driving Score on Bench2Drive by routing vision-language tokens to semantic experts while handling trajectory computation separately, demonstrating advances in multimodal AI for real-world driving tasks.

VECTOR-Drive addresses a fundamental challenge in end-to-end autonomous driving: how to effectively combine semantic understanding from vision-language models with precise motion planning without sacrificing performance in either domain. The research represents a meaningful step forward in multimodal AI architecture design, introducing a hybrid approach that maintains token coupling through shared attention layers while strategically separating feed-forward network computation by task type. This selective decoupling reduces computational conflicts between language reasoning and trajectory prediction—two inherently different objectives that compete for model capacity.

The broader context shows autonomous driving research increasingly converging on vision-language foundations. Rather than building specialized perception modules, researchers now leverage semantic priors learned from large-scale pretraining. VECTOR-Drive's innovation lies in its routing mechanism: vision and language tokens flow to dedicated Vision-Language Experts, while motion-related tokens route to Trajectory Experts. This design respects the asymmetric nature of the problem—semantic understanding benefits from joint processing, while action generation requires specialized computation pathways.

The practical implications extend across autonomous vehicle development and multimodal AI research. Strong benchmark performance (88.91 Driving Score, outperforming baselines) validates that selective decoupling can outperform both fully shared and fully separated architectures. For the autonomous driving industry, this suggests that future systems may benefit from architecture-level specialization rather than generic end-to-end approaches. The flow-matching planner for action token refinement also indicates progress in translating discrete model outputs into smooth, executable motion plans—a critical gap in deploying learned driving policies.

• →VECTOR-Drive achieves state-of-the-art 88.91 Driving Score through tightly coupled vision-language and trajectory expert routing
• →Shared self-attention with task-specific feed-forward networks reduces computational conflict while preserving semantic-motion coupling
• →The architecture balances fully shared backbones against decoupled pipelines, advancing the optimal design space for multimodal autonomous driving
• →Flow-matching planner refines noisy action tokens into executable waypoints and speed profiles, bridging discrete predictions to continuous control
• →Results validate that selective architectural decoupling outperforms both fully coupled and fully separated VLA approaches for driving tasks