Boosting Efficient Reinforcement Learning for Vision-and-Language Navigation with Open-Sourced LLM

Vision-and-Language Navigation (VLN) requires an agent to navigate in photo-realistic environments based on language instructions. Existing methods typically employ imitation learning to train agents. However, approaches based on recurrent neural networks suffer from poor generalization, while transformer-based methods are too large in scale for practical deployment. In contrast, reinforcement learning (RL) agents can overcome dataset limitations and learn navigation policies that adapt to environment changes. However, without expert trajectories for supervision, agents struggle to learn effective long-term navigation policies from sparse environment rewards. Instruction decomposition enables agents to learn value estimation faster, making agents more efficient in learning VLN tasks. We propose the Decomposing Instructions with Large Language Models for Vision-and-Language Navigation (DILLM-VLN) method, which decomposes complex navigation instructions into simple, interpretable sub-instructions using a lightweight, open-sourced LLM and trains RL agents to complete these sub-instructions sequentially. Based on these interpretable sub-instructions, we introduce the cascaded multi-scale attention (CMA) and a novel multi-modal fusion discriminator (MFD). CMA integrates instruction features at different scales to provide precise textual guidance. MFD combines scene, object, and action information to comprehensively assess the completion of sub-instructions. Experiment results show that DILLM-VLN significantly improves baseline performance, demonstrating its potential for practical applications.