Curiosity-Driven Zero-Shot Object Navigation With Vision-Language Models

Navigating toward specific objects in unknown environments, known as Zero-Shot Object Navigation, is a critical aspect of Embodied Intelligence. Current learning-based and modular methods struggle with inadequate exploration strategies and inefficient decision-making, limiting their effectiveness. To address these challenges, we propose a novel framework leveraging the reasoning capabilities of vision-language models (VLMs) to emulate human-like assessments of environmental suitability. These assessments are quantified into a dynamic value map that guides the agent toward promising regions while penalizing revisits to redundant areas, enabling adaptive and efficient exploration. A key innovation of our approach lies in integrating human-inspired reasoning directly into the navigation loop, ensuring that agents can distinguish between high-priority and low-priority environments at each step. Evaluated on the HM3D dataset, our method achieves an absolute improvement of 12.1% (from 56.5% to 68.6%) in success rate over state-of-the-art approaches. Further qualitative analysis demonstrates that balancing exploration with curiosity significantly enhances navigation performance, showcasing the potential of our framework for complex real-world tasks.