Horizon-Adaptive Reinforcement Learning for Vision-Language Navigation in Unseen Environments

Abstract

Vision-and-language navigation represents a fundamental challenge in the development of embodied artificial intelligence, requiring an agent to interpret unconstrained natural language instructions and execute sequential navigational actions within complex, photorealistic environments. Despite significant advancements in cross-modal representation learning and navigation strategies, existing models frequently struggle to generalize to unseen environments. A primary bottleneck lies in the rigid, predefined temporal horizons employed by standard reinforcement learning frameworks, which force agents into either premature termination or persistent wandering when faced with environmental ambiguities. To address this limitation, this paper introduces a horizon-adaptive reinforcement learning framework designed to dynamically modulate the temporal bounds of an episode based on the real-time alignment between the agent trajectory and the linguistic instruction. By continuously assessing the semantic progress of the navigation task, the proposed architecture optimally expands or contracts the action horizon, thereby encouraging efficient exploration while mitigating catastrophic error accumulation in unfamiliar topologies. Comprehensive evaluations conducted on standard benchmark datasets demonstrate that the horizon-adaptive approach substantially reduces the generalization gap between seen and unseen environments. The framework achieves superior performance across primary navigation metrics, particularly in the success rate weighted by path length, indicating a more efficient and human-like navigation paradigm. The findings suggest that dynamic temporal constraints are essential for robust decision-making in embodied agents.