LLM-Guided Agentic Floor Plan Parsing for Accessible Indoor Navigation of Blind and Low-Vision People
Indoor navigation continues to pose significant challenges for blind and low-vision (BLV) individuals, primarily due to the reliance on expensive and complex infrastructure that varies from building to building. A new study, recently published on arXiv (2604.23970v1), proposes an innovative solution that leverages an agentic framework to transform a simple floor plan image into a structured knowledge base, facilitating safe and accessible navigation instructions with minimal infrastructure requirements.
The Proposed Framework
The proposed system operates in two distinct phases:
- Multi-Agent Module: This module is responsible for parsing the floor plan into a spatial knowledge graph. It employs a self-correcting pipeline featuring iterative retry loops and corrective feedback to ensure accuracy.
- Path Planner: Once the knowledge graph is established, the Path Planner generates navigation instructions, while a Safety Evaluator agent assesses potential hazards along each proposed route.
This two-phase approach not only enhances the precision of navigation instructions but also provides a scalable solution that can be adapted to various indoor environments without the need for extensive modifications or installations.
Performance Evaluation
The system was rigorously tested in real-world settings, specifically in the UMBC Math and Psychology building, across two floors (MP-1 and MP-3). The evaluation included comparisons with existing methodologies, particularly focusing on the capabilities of the strongest single-call baseline, Claude 3.7 Sonnet.
Results Summary
On the MP-1 floor, the new system achieved remarkable success rates:
- 92.31% for short routes
- 76.92% for medium routes
- 61.54% for long routes
These results significantly outperformed the Claude 3.7 Sonnet baseline, which recorded success rates of 84.62%, 69.23%, and 53.85% for the corresponding route lengths.
Similarly, on the MP-3 floor, the new framework delivered the following success rates:
- 76.92% for short routes
- 61.54% for medium routes
- 38.46% for long routes
In this instance, the strongest baseline recorded success rates of 61.54%, 46.15%, and 23.08% for the same categories, underscoring the effectiveness of the proposed system.
Conclusion
The findings from this study highlight the potential of LLM-guided agentic frameworks in enhancing indoor navigation for BLV individuals. The consistent performance improvements over single-call LLM baselines demonstrate not only the viability of this approach but also its scalability, offering a promising avenue for future research and development in accessible technology.
As urban spaces continue to evolve, the integration of such innovative solutions can play a pivotal role in fostering inclusivity and ensuring that all individuals, regardless of visual impairment, can navigate their environments safely and independently.
Related AI Insights
- Agentic AI for Autonomous Protein-Protein Interaction Analysis
- EU-AI-Act Compliant Time-Series Forecasting Package
- FinGround: Reducing Financial AI Errors with Claim Verification
- ClawTrace: Cost-Aware Tracing for Efficient LLM Skill Distillation
- Can We Trust AI in Scientific Peer Review?
- AI Information-Theoretic Measures: Practical Selection Guide
- Vibe Medicine: Human-AI Collaboration in Biomedical Research
- Tandem: Efficient Reasoning with Large & Small Language Models
- IndustryAssetEQA: AI for Smarter Industrial Asset Maintenance
- MetaGAI: Benchmark for Generative AI Model & Data Cards
