Explaining Navigation Decisions: Stage-Aware Visual Attribution for Vision-Language Navigation

TL;DR

This paper introduces a stage-aware attribution framework for Vision-Language Navigation that addresses instruction-stage insensitivity by modeling temporal visual-linguistic attributions across navigation stages. It uses instruction-space projections and spatial filtering to refine explanations with relevant historical evidence, improving attribution quality and enabling failure analysis.

Why It Matters

Explainable AI is critical for trustworthy robot navigation in real-world settings. This method enhances model interpretability by aligning explanations with task stages, aiding debugging and deployment in complex environments like those encountered by physical robots.

Key Findings

• Existing explainability methods are stage-insensitive, producing inconsistent explanations across navigation phases.
• The proposed stage-aware framework improves attribution quality by aggregating temporally relevant visual evidence guided by instruction stage estimation.
• The method serves as a diagnostic tool and enables reliable real-world navigation on a Unitree G1 robot.