Constraint-Aware Zero-Shot Vision-Language Navigation in Continuous Environments.

Area of Science:

• Robotics
• Artificial Intelligence
• Computer Vision

Background:

• Vision-Language Navigation in Continuous Environments (VLN-CE) is challenging, especially in zero-shot settings due to limited training data and environmental structure.
• Existing methods struggle with navigation stability and adaptability without expert demonstrations.

Purpose of the Study:

• To develop a novel approach for zero-shot VLN-CE that overcomes limitations of current methods.
• To enhance navigation performance and stability in unseen environments.

Main Methods:

• Propose Constraint-Aware Navigator (CA-Nav), reframing zero-shot VLN-CE as sequential, constraint-aware sub-instruction completion.
• Utilize Constraint-Aware Sub-instruction Manager (CSM) to define completion criteria and track progress.
• Employ Constraint-Aware Value Mapper (CVM) with superpixel clustering for on-the-fly value map generation and refinement.

Main Results:

• CA-Nav achieves state-of-the-art performance on R2R-CE and RxR-CE benchmarks.
• Surpassed previous best methods by 12% and 13% in Success Rate on validation unseen splits.
• Demonstrated effectiveness in real-world robot deployments across diverse indoor scenes.

Conclusions:

• CA-Nav offers a robust and effective solution for zero-shot VLN-CE.
• The proposed constraint-aware framework significantly advances navigation capabilities in complex environments.
• CA-Nav shows strong potential for practical applications in autonomous navigation.