OmniV2X: A Generative Foundation Planner for Efficient End-to-End Cooperative Driving

OmniV2X is a generative foundation model that enables vehicle-to-everything (V2X) cooperative driving by processing multi-modal, multi-agent data without requiring dense 3D perception or shared representations. The model achieves state-of-the-art performance on the DAIR-V2X-Seq dataset while using 90% less fine-tuning data and consuming less than 1% of typical communication bandwidth.

OmniV2X addresses fundamental inefficiencies in autonomous vehicle cooperation by introducing a generative planning approach that bypasses traditional sensor fusion bottlenecks. Rather than converting multi-modal sensor inputs into unified 3D representations—a computationally expensive process vulnerable to data scarcity—the model processes independent context sequences directly, leveraging cross-attention mechanisms to extract relevant information dynamically. This architectural choice represents a meaningful shift in how cooperative driving systems handle heterogeneous data sources.

The autonomous vehicle industry has struggled with the scalability and standardization challenges inherent in V2X communication. Existing fusion-based approaches often require extensive fine-tuning on cooperative datasets and demand high communication bandwidth to exchange processed sensor representations. OmniV2X's foundation model approach, pre-trained on large-scale single-agent planning data, enables efficient transfer learning to cooperative scenarios through lightweight V2X tokens that comply with emerging communication standards.

From a deployment perspective, the practical advantages are substantial. The model's ability to achieve superior performance using only 10% of typical fine-tuning data accelerates commercialization timelines and reduces costly data collection requirements. The 99% reduction in communication bandwidth has direct implications for real-world V2X infrastructure, enabling broader adoption across vehicle fleets with heterogeneous communication capabilities.

The comprehensive evaluation on standardized benchmarks validates the approach's robustness under real-world constraints, signaling maturity beyond prototype stage. As autonomous vehicles move toward widespread deployment, computational efficiency and communication efficiency increasingly determine commercial viability. OmniV2X's advances in both dimensions position generative modeling as a competitive alternative to traditional perception pipelines in cooperative driving scenarios.

• →OmniV2X uses generative foundation models to achieve state-of-the-art V2X cooperative driving with 90% less fine-tuning data than existing methods.
• →The model processes multi-modal sensor data independently via cross-attention rather than fusing into shared representations, reducing computational overhead.
• →Communication bandwidth requirements are reduced by over 99% through lightweight, standards-compliant V2X tokens.
• →Pre-training on single-agent planning datasets enables efficient transfer to cooperative scenarios, improving scalability.
• →Foundation model approach demonstrates robustness under real-world constraints while maintaining practical deployment feasibility.