PriorNav: Prior Knowledge Enhanced Zero-Shot Goal Navigation via Multi-Step Iterative Reasoning.

Area of Science:

• Robotics
• Artificial Intelligence
• Computer Vision

Background:

• Zero-shot goal navigation demands robust scene understanding and reasoning for locating targets in novel environments.
• Current methods often fall short due to limited external knowledge integration and weak multi-step reasoning capabilities.

Purpose of the Study:

• To introduce PriorNav, a novel framework designed to enhance zero-shot goal navigation through prior-knowledge integration.
• To improve agent performance in complex navigation tasks by leveraging semantic, instance, and relational knowledge.

Main Methods:

• PriorNav constructs a unified retrievable knowledge space from diverse prior knowledge types.
• It maintains a knowledge-enhanced scene graph by fusing retrieved priors with real-time observations.
• The framework employs progressive decision-making via multi-step iterative reasoning across exploration, verification, and approach stages.

Main Results:

• PriorNav demonstrated significant improvements in success rates: 3.5% for Object-Goal, 13.3% for Image-Instance Goal, and 3.5% for Text-Goal navigation.
• The proposed method outperformed existing training-free baselines across all tested navigation tasks.
• Ablation studies confirmed the efficacy of multi-level prior knowledge, scene-graph enhancement, and iterative reasoning.

Conclusions:

• Integrating prior knowledge with explicit reasoning offers a promising avenue for advancing zero-shot goal navigation.
• PriorNav establishes a new state-of-the-art by effectively combining external knowledge with robust reasoning mechanisms.