LaF-GRPO: In-Situ Navigation Instruction Generation for the Visually Impaired via GRPO with LLM-as-Follower Reward

The recent study on LaF-GRPO represents a pivotal advancement in assistive technology for visually impaired individuals by focusing on generating precise navigation instructions. By employing a Vision-Language Model (VLM) that simulates user responses, LaF-GRPO enhances the accuracy of these instructions while significantly reducing the need for costly real-world data collection. The introduction of the NIG4VI dataset, comprising 27,000 samples, provides a robust resource for training and evaluation, addressing the current scarcity of dedicated benchmarks in this field. Experimental results indicate that LaF-GRPO improves instruction accuracy, as evidenced by a 14% boost in BLEU scores and superior METEOR scores compared to existing models. This research not only contributes to the development of more intuitive navigation aids but also underscores the importance of innovative methodologies in creating accessible solutions for the visually impaired community.