OmniDrive-R1 enhances autonomous driving VLMs with reinforcement-driven visual grounding

Researchers have introduced OmniDrive-R1, a novel framework for autonomous driving that integrates perception and reasoning using an interleaved Multi-modal Chain-of-Thought (iMCoT) mechanism. This approach addresses object hallucination issues common in Vision-Language Models by employing a reinforcement-driven visual grounding capability. The system utilizes a unique annotation-free training pipeline with the Clip-GRPO algorithm, which generates a grounding reward without requiring dense localization labels. Experiments show OmniDrive-R1 significantly boosts reasoning scores and accuracy compared to baseline models. AI

Summary written by gemini-2.5-flash-lite from 1 source. How we write summaries →

IMPACT Introduces a novel approach to improve VLM reliability in safety-critical autonomous driving applications.

RANK_REASON This is a research paper detailing a new model and methodology for autonomous driving.

Read on arXiv cs.AI →

• OmniDrive-R1
• Vision-Language Models
• Chain-of-Thought
• Clip-GRPO
• Qwen2.5VL-7B
• DriveLMM-o1
• Zhenguo Zhang

• paper
• model release