iDriveVLA: Advanced Planning Framework for Autonomous Driving Systems

Paper: Evaluation Is All You Need for Multi-Modal Autonomous Driving

Authors: Zeyu He, Shiqi Liu, Ke Chen, Yun Yan, Jinzi Wu, Dianqiao Lei, Sirui Wang, ShuRui Peng, Tao Chen, Zhuo Huang, Yu Wu, Yadong Shao, Zhichao Li, Ke Sun, Yang Guan, Keqiang Li, Shengbo Eben Li

Executive Summary

The research introduces iDriveVLA, an advanced planning framework designed to enhance autonomous driving systems. As autonomous vehicles navigate complex environments, they must account for various potential behaviors in uncertain scenarios. Current methodologies primarily focus on improving the variety of potential trajectories and the representation of these paths. However, they often struggle with a key issue: while they may generate a wide array of possible actions, they do not consistently select the safest or most effective option. This gap indicates that there is significant potential for better decision-making that has not yet been fully realized.

iDriveVLA aims to bridge this gap by refining how autonomous systems evaluate and select trajectories. The framework incorporates two main components: a Safety-aware Scorer and a VLM-guided Modulator. The Safety-aware Scorer evaluates the quality and risk of various trajectory options, helping the system identify which paths are safest under specific conditions. The VLM-guided Modulator adjusts the evaluation criteria based on the context of the scene, ensuring that the assessment aligns with real-world driving challenges.

To optimize the performance of iDriveVLA, the researchers developed a progressive training strategy. This strategy consists of three key steps: first, imitating successful candidate trajectories to build a strong foundational model; second, refining the space of possible candidate trajectories to improve overall quality; and third, aligning the evaluation process with a high-performing oracle, which represents an ideal decision-making model.

The effectiveness of iDriveVLA was demonstrated through benchmarks on the NAVSIM v1 leaderboard, where it achieved a performance score of 94.95 on the PDMS metric, surpassing previous results established by human experts. This performance indicates not only that iDriveVLA can generate high-quality trajectories but also that it can select the best ones in a more reliable manner.

The implications of this research are significant for the development of autonomous driving technologies. By improving the decision-making process in complex driving environments, iDriveVLA could enhance the safety and efficiency of autonomous vehicles. As companies in the automotive sector, such as Tesla, Waymo, or traditional automakers venturing into autonomous technologies, seek to refine their systems, frameworks like iDriveVLA may provide valuable insights and methodologies to foster more intelligent and context-aware driving behavior.

Academic Abstract

Multi-modal planning is promising for autonomous driving by representing multiple plausible behaviors in ambiguous and long-tail scenarios. Existing methods mainly focus on improving trajectory multi-modality, enhancing trajectory representations, or reshaping the candidate distribution. Nevertheless, we identify a pronounced generation-evaluation asymmetry in multi-modal planning: despite strong oracle performance, existing planners often fail to reliably select the best available candidate, leaving substantial planning potential unrealized. To address this challenge, we propose iDriveVLA, a multi-modal planning framework that improves the candidate trajectory space while enabling more reliable and context-aware trajectory evaluation. Specifically, iDriveVLA introduces a unified trajectory evaluator comprising a Safety-aware Scorer for quality and risk estimation, together with a VLM-guided Modulator for scene-adaptive criterion weighting. We further develop an oracle-aligned progressive training strategy consisting of candidate imitation pretraining, candidate space refinement, and semantic ranking alignment. On the public NAVSIM v1 leaderboard, iDriveVLA achieves a new state-of-the-art performance of 94.95 PDMS, surpassing the human-expert reference.