Search papers, labs, and topics across Lattice.
UniDriveVLA, a unified driving VLA model, is proposed to address the dilemma between spatial perception and semantic reasoning in autonomous driving by decoupling these functionalities into separate experts: driving understanding, scene perception, and action planning. This is achieved through a Mixture-of-Transformers architecture coordinated by masked joint attention and a sparse perception paradigm. Experiments on nuScenes and Bench2Drive demonstrate state-of-the-art performance in both open-loop and closed-loop evaluations, showcasing strong performance across perception, prediction, and understanding tasks.
Autonomous driving models no longer need to compromise between spatial perception and semantic reasoning: UniDriveVLA's expert decoupling unlocks state-of-the-art performance across a range of driving tasks.
Vision-Language-Action (VLA) models have recently emerged in autonomous driving, with the promise of leveraging rich world knowledge to improve the cognitive capabilities of driving systems. However, adapting such models for driving tasks currently faces a critical dilemma between spatial perception and semantic reasoning. Consequently, existing VLA systems are forced into suboptimal compromises: directly adopting 2D Vision-Language Models yields limited spatial perception, whereas enhancing them with 3D spatial representations often impairs the native reasoning capacity of VLMs. We argue that this dilemma largely stems from the coupled optimization of spatial perception and semantic reasoning within shared model parameters. To overcome this, we propose UniDriveVLA, a Unified Driving Vision-Language-Action model based on Mixture-of-Transformers that addresses the perception-reasoning conflict via expert decoupling. Specifically, it comprises three experts for driving understanding, scene perception, and action planning, which are coordinated through masked joint attention. In addition, we combine a sparse perception paradigm with a three-stage progressive training strategy to improve spatial perception while maintaining semantic reasoning capability. Extensive experiments show that UniDriveVLA achieves state-of-the-art performance in open-loop evaluation on nuScenes and closed-loop evaluation on Bench2Drive. Moreover, it demonstrates strong performance across a broad range of perception, prediction, and understanding tasks, including 3D detection, online mapping, motion forecasting, and driving-oriented VQA, highlighting its broad applicability as a unified model for autonomous driving. Code and model have been released at https://github.com/xiaomi-research/unidrivevla
