Rule-VLN: New AI Benchmark for Socially Compliant Navigation in Urban Environments
A new guideline called Rule-VLN has been launched to improve the flaws found in current Vision-and-Language Navigation systems. These systems often focus too much on physical layout while overlooking crucial regulatory aspects. The new urban benchmark consists of 29,000 nodes and includes 177 unique regulatory categories across 8,000 restricted nodes. It challenges agents with specific visual and behavioral constraints through four levels of difficulty. To boost safety awareness in pre-trained agents, a tool called the Semantic Navigation Rectification Module (SNRM) has been developed. This universal solution merges a visual perception framework with an epistemic mental map. This study, which explores the evolution of AI from basic task completion to adhering to social norms, was shared on arXiv under the ID 2604.16993v1.
Key facts
- Rule-VLN is the first large-scale urban benchmark for rule-compliant navigation
- The environment spans 29,000 nodes with 8,000 constrained nodes
- 177 diverse regulatory categories are incorporated across the benchmark
- Four curriculum levels challenge agents with fine-grained constraints
- The Semantic Navigation Rectification Module (SNRM) is a universal, zero-shot module
- SNRM integrates visual perception VLM framework with epistemic mental map
- Current VLN agents suffer from a "goal-driven trap" prioritizing physical over semantic rules
- Research announced on arXiv under identifier 2604.16993v1
Entities
Institutions
- arXiv