Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

DiffVLA++: Bridging Cognitive Reasoning and End-to-End Driving through Metric-Guided Alignment

Yu Gao, Anqing Jiang, Yiru Wang, Wang Jijun, Hao Jiang, Zhigang Sun, Heng Yuwen, Wang Shuo, Hao Zhao, Sun Hao

TL;DR

DiffVLA++ addresses the mismatch between world-knowledge-driven cognitive reasoning and physically feasible planning in autonomous driving by integrating a differentiable Vision-Language-Action (VLA) module, a dense BEV end-to-end (E2E) module, and a metric-guided trajectory scorer. A shared BEV feature space and an eight-metric scorer enable principled alignment between semantically rich VLA trajectories and physically grounded E2E trajectories, using an 8192-trajectory vocabulary and joint training on Navsim data. Post-processing with a drivable-area filter and a weighted metric ranking enables robust selection, with offline ensemble choosing the final output. On the Navhard two-stage test and ICCV 2025 leaderboard, DiffVLA++ achieves an EPDMS of 49.12, surpassing individual branches and demonstrating improved reliability in long-tail scenarios.

Abstract

Conventional end-to-end (E2E) driving models are effective at generating physically plausible trajectories, but often fail to generalize to long-tail scenarios due to the lack of essential world knowledge to understand and reason about surrounding environments. In contrast, Vision-Language-Action (VLA) models leverage world knowledge to handle challenging cases, but their limited 3D reasoning capability can lead to physically infeasible actions. In this work we introduce DiffVLA++, an enhanced autonomous driving framework that explicitly bridges cognitive reasoning and E2E planning through metric-guided alignment. First, we build a VLA module directly generating semantically grounded driving trajectories. Second, we design an E2E module with a dense trajectory vocabulary that ensures physical feasibility. Third, and most critically, we introduce a metric-guided trajectory scorer that guides and aligns the outputs of the VLA and E2E modules, thereby integrating their complementary strengths. The experiment on the ICCV 2025 Autonomous Grand Challenge leaderboard shows that DiffVLA++ achieves EPDMS of 49.12.

DiffVLA++: Bridging Cognitive Reasoning and End-to-End Driving through Metric-Guided Alignment

TL;DR

DiffVLA++ addresses the mismatch between world-knowledge-driven cognitive reasoning and physically feasible planning in autonomous driving by integrating a differentiable Vision-Language-Action (VLA) module, a dense BEV end-to-end (E2E) module, and a metric-guided trajectory scorer. A shared BEV feature space and an eight-metric scorer enable principled alignment between semantically rich VLA trajectories and physically grounded E2E trajectories, using an 8192-trajectory vocabulary and joint training on Navsim data. Post-processing with a drivable-area filter and a weighted metric ranking enables robust selection, with offline ensemble choosing the final output. On the Navhard two-stage test and ICCV 2025 leaderboard, DiffVLA++ achieves an EPDMS of 49.12, surpassing individual branches and demonstrating improved reliability in long-tail scenarios.

Abstract

Conventional end-to-end (E2E) driving models are effective at generating physically plausible trajectories, but often fail to generalize to long-tail scenarios due to the lack of essential world knowledge to understand and reason about surrounding environments. In contrast, Vision-Language-Action (VLA) models leverage world knowledge to handle challenging cases, but their limited 3D reasoning capability can lead to physically infeasible actions. In this work we introduce DiffVLA++, an enhanced autonomous driving framework that explicitly bridges cognitive reasoning and E2E planning through metric-guided alignment. First, we build a VLA module directly generating semantically grounded driving trajectories. Second, we design an E2E module with a dense trajectory vocabulary that ensures physical feasibility. Third, and most critically, we introduce a metric-guided trajectory scorer that guides and aligns the outputs of the VLA and E2E modules, thereby integrating their complementary strengths. The experiment on the ICCV 2025 Autonomous Grand Challenge leaderboard shows that DiffVLA++ achieves EPDMS of 49.12.
Paper Structure (11 sections, 5 equations, 1 figure, 2 tables)

This paper contains 11 sections, 5 equations, 1 figure, 2 tables.

Table of Contents

  1. Introduction
  2. Fully-Differentiable VLA Model for E2E Driving
  3. Conventional End-to-End Driving Model
  4. Trajectory Planning Head
  5. Metric-Guided Alignment
  6. Post-Processing
  7. Experiments
  8. Training for VLA
  9. Training for E2E and scorer in Metric-Guided Alignment
  10. Experiments result
  11. Conclusion

Figures (1)

  • Figure 1: Overview of the DiffVLA++ architecture. It consists of three main components: (i) a VLA module, (ii) a conventional E2E module, both capable of directly generating planning trajectories, and (iii) a trajectory scorer that serves as a metric-guided aligner to unify their outputs.