Embodied intelligence requires an agent to complete autonomous navigation in continuous environments following natural language instructions. Existing methods typically output mid-level natural language actions or low-level control commands. However, there is a lack of effective coordination between high-level reasoning and continuous control, which leaves room for improvement in decision interpretability and control stability. To address these issues, we propose the VLN-COT framework, which integrates vision-language navigation models with explicit chain-of-thought reasoning. This framework introduces a generative reasoning mechanism to enhance decision interpretability through structured reasoning processes. It also adopts a dual-modal action representation that simultaneously predicts mid-level natural language actions and low-level continuous control commands, achieving joint optimization between high-level reasoning and low-level execution. During training, we combine supervised fine-tuning with GRPO-based reinforcement learning and design three complementary reward functions to obtain a stable and interpretable navigation decisions. Experimental results on the VLN-CE benchmark show that VLN-COT achieves significant improvements in key metrics such as success rate and path efficiency, validating the effectiveness of explicit reasoning and dual-modal action representation in continuous vision-language navigation.
原文链接
分享
引用
