BROWSE
Detail
Conference Paper
구조화된 RLAIF를 활용한 보행 상황 설명 VLM의 성능 개선
Cited - time in 
Share 
Share
- Issue Date
- 2026-07
- Citation
- 제어·로봇·시스템학회 학술 대회 (ICROS) 2026, pp.895-896
- Publisher
- 제어·로봇·시스템학회
- Language
- Korean
- Type
- Conference Paper
- Abstract
- Assistive navigation for visually impaired individuals increasingly relies on vision-language models (VLMs), but existing methods require costly human annotations and still show limitations in depth and contextual understanding. We propose a structured reinforcement learning from AI feedback (RLAIF) framework that decomposes feedback into multiple attributes such as walkability, distance, direction, and contextual awareness. Using LLM-based structured rewards and diverse outputs from baseline and off-the-shelf VLMs, our approach enables fine-grained model improvement without additional human annotations.
- Keyword
- Vision-Language Models (VLMs), Assistive Navigation, RLAIF, Structured Feedback.
- KSP Keywords
- Contextual awareness, Fine grained(FG), Language Models, Model improvement, Multiple attributes, Off-the-shelf, Visually impaired individuals, assistive navigation, reinforcement learning(RL)