VLAN benchmark tests whether AI models navigate for different bodies, not just different scenes
Researchers introduce Vision-Language Accessible Navigation, a benchmark of 1,000 scenes and 20,000 trajectories with embodiment-specific accessibility labels, finding current vision-language models struggle to adapt safety judgments across user profiles.
In plain words
Whether a path is walkable depends on who or what is moving through it: a route fine for a wheeled robot may not suit a person with a mobility impairment. The VLAN benchmark provides navigation data tagged to specific agents, including robots and people with visual or mobility impairments. Its creators report that today's vision-language models often fail to adjust obstacle relevance and safety criteria to match the user.
What was announced
A team presenting the Vision-Language Accessible Navigation (VLAN) benchmark argues that accessible mobility is user-conditioned rather than a property of the scene itself, since a path passable for one embodiment can be inaccessible for another. Existing benchmarks, the authors say, evaluate traversability under a fixed and limited set of embodiments. VLAN instead contains 1,000 scenes, 20,000 navigation trajectories, and 160,000 keyframes spanning varied lighting, environmental conditions, scene semantics, and path geometries. Each trajectory carries embodiment-specific actions and accessibility features for legged and wheeled agents, including humanoid and quadruped robots and people with visual or mobility impairments.
Everyday obstacles and accessibility labels fall into four categories: Passable, Bypass, HighRisk, and Blocked. Rather than judging success by reaching a goal alone, the benchmark scores physical safety and alignment with user-specific constraints through goal-reaching, collision-aware, and accessibility-aligned metrics. When the authors ran modern vision-language models through it, they found the models often failed to consistently adapt obstacle relevance, path selection, and safety criteria across user profiles. They conclude that accessibility-aware navigation needs dedicated datasets, tasks, and metrics beyond conventional embodied navigation evaluation. The data and code are open-source on GitHub.
What it does not show
- The evidence is the paper's abstract only; the full methodology, experimental setup, and detailed results are not covered here.
- The finding that models fail to adapt across user profiles reflects the authors' own benchmark evaluation and has not been independently confirmed.
- The abstract does not specify which vision-language models were tested or quantify how often they failed.
Sources
Source-support review completed 17 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 17 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.