Researchers have introduced VLAN, a new benchmark dataset designed to evaluate vision-language models for accessible navigation. VLAN includes 1,000 scenes and 20,000 navigation trajectories, with specific actions and accessibility features tailored for various embodiments like legged and wheeled robots, as well as people with visual or mobility impairments. The dataset categorizes obstacles as Passable, Bypass, HighRisk, and Blocked, and evaluates navigation based on physical safety and user-specific constraints. Current vision-language models struggle to adapt obstacle relevance and safety criteria across different user profiles, highlighting the need for dedicated datasets and metrics for accessibility-aware navigation. AI
IMPACT This benchmark could drive the development of more inclusive AI navigation systems for robots and assistive technologies.
RANK_REASON The item describes a new benchmark dataset and associated research paper published on arXiv. [lever_c_demoted from research: ic=1 ai=1.0]
- arXiv
- Humanoid Robots
- legged agents
- people with visual impairments
- Quadruped Robots
- vision-language model
- VLAN
- wheeled agents
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →