Vision-Language Navigation With Beam-Constrained Global Normalization.

Area of Science:

• Artificial Intelligence
• Robotics
• Computer Vision

Background:

• Vision-language navigation (VLN) agents interpret natural language instructions to move within environments.
• Sequence-to-sequence models are effective for VLN, achieving state-of-the-art results.
• Beam-search decoding improves VLN performance by evaluating entire navigation trajectories.

Purpose of the Study:

• To address the bias in trajectory-level scores used in beam-search decoding for VLN.
• To propose a global normalization strategy for more accurate trajectory ranking.
• To enhance the performance of agents in vision-language navigation tasks.

Main Methods:

• Developed a global normalization strategy to rescale trajectory scores.
• Introduced two novel global score functions for reranking candidate trajectories.
• Implemented and tested the proposed method on the Room-to-Room (R2R) dataset.

Main Results:

• The global normalization strategy significantly alleviates bias in trajectory scoring.
• Experimental results demonstrate superior performance compared to baseline methods.
• The proposed approach achieves competitive results on the VLN leaderboard.

Conclusions:

• Global score normalization is an effective technique for improving VLN.
• The method enhances the reliability and accuracy of trajectory evaluation in navigation tasks.
• This work contributes to advancing the capabilities of embodied AI agents.