EvolveNav: Empowering LLM-Based Vision-Language Navigation via Self-Improving Embodied Reasoning

Area of Science:

• Artificial Intelligence
• Robotics
• Computer Vision

Background:

• Large Language Models (LLMs) show promise for Vision-Language Navigation (VLN).
• Current LLM-based VLN methods struggle with explainability and domain gaps.
• Chain-of-Thought (CoT) training improves accuracy and interpretability but faces challenges with label availability and overfitting in complex navigation tasks.

Purpose of the Study:

• To introduce EvolveNav, a self-improving embodied reasoning paradigm for LLM-based VLN.
• To enhance navigational decision-making accuracy and interpretability in LLM-driven agents.
• To develop adaptable and generalizable reasoning capabilities for embodied AI.

Main Methods:

• EvolveNav employs a two-stage training process: Formalized CoT Supervised Fine-Tuning and Self-Reflective Post-Training.
• The first stage uses curated CoT labels to activate reasoning and increase speed.
• The second stage uses self-generated reasoning outputs as enriched labels, with an auxiliary task to refine correct reasoning patterns.

Main Results:

• EvolveNav demonstrates consistent superiority over existing LLM-based VLN approaches across multiple benchmarks (R2R, REVERIE, CVDN, SOON).
• The approach shows effectiveness in both task-specific and cross-task training paradigms.
• EvolveNav improves navigational reasoning adaptability and generalizability.

Conclusions:

• EvolveNav presents a novel self-improving paradigm for embodied reasoning in LLM-based VLN.
• The method effectively addresses limitations of previous approaches, enhancing both performance and interpretability.
• EvolveNav paves the way for self-evolving AI agents in embodied AI research.