MOJITO: Modal Joint Learning for Unified End-to-End Autonomous Driving

2026-07-26

AI Summary

This work addresses the performance limitations of conventional end-to-end autonomous driving systems, which suffer from multimodal information loss due to their perception-planning cascaded architecture. To overcome this, the authors propose MOJITO, a unified sensor-to-action framework that eliminates cascaded design and instead introduces a chunked modality-joint attention mechanism within a Transformer architecture. This enables synchronous fusion of image, LiDAR, and action features, facilitating joint optimization of multimodal representation learning and action prediction. MOJITO supports deep integration with vision foundation models, exhibits strong instruction-following capability, and demonstrates excellent scalability. Evaluated on NAVSIM v1 and v2, it achieves state-of-the-art results with 88.9 PDMS and 88.4 EPDMS, respectively, while also showing superior trajectory diversity and generalization performance.

Abstract

End-to-end autonomous driving systems commonly follow a cascaded two-stage pipeline where a perception stage compresses multi-modal sensor inputs into a compact context and a downstream planner predicts trajectories conditioned on this context. We argue that this one-way perception-to-planning interface forces sensor inputs into a compact representation, losing the fine-grained details critical for planning. Moreover, by constraining the planner to this compressed context, it is difficult to leverage the rich representations offered by modern vision foundation models. To address these issues, we propose MOJITO, a unified sensor-to-action framework for end-to-end autonomous driving built on modal joint learning. MOJITO removes the cascaded interface and instead performs block-wise Modal Joint Attention that simultaneously updates action, image, and LiDAR features, allowing the planner to directly access multi-modal features during action generation. MOJITO achieves 88.9 PDMS on the NAVSIM v1 dataset and 88.4 EPDMS on the more challenging NAVSIM v2 dataset, setting a new state-of-the-art. Extensive experiments further demonstrate strong scalability, instruction following, and diverse trajectory generation. Code and models are available at https://github.com/mumucc01/MOJITO.

Problem

• end-to-end autonomous driving
• multi-modal perception
• planning interface
• feature compression
• vision foundation models

Innovation

• Modal Joint Learning
• End-to-End Autonomous Driving
• Multi-Modal Fusion
• Joint Attention
• Sensor-to-Action Framework