What they found
The authors built VLAN, a dataset of 1,000 scenes, 20K navigation trajectories, and 160K keyframes. It covers varied lighting, environments, scene types, and path shapes. Each trajectory comes with actions and accessibility features for legged and wheeled agents, including humanoid and quadruped robots and people with visual or mobility impairments. Obstacles and accessibility labels are sorted into four classes: Passable, Bypass, HighRisk, and Blocked. The authors tested modern vision-language models and found they often fail to adapt obstacle relevance, path selection, and safety criteria consistently across user profiles. The abstract gives no numbers for how large those failures are.
Why it matters
A path that one user can walk may be closed to another, but existing benchmarks usually test only one fixed body type. VLAN scores more than whether the goal is reached. It also scores collisions and how well the route fits a user's own constraints. This points to a gap that matters for SI systems meant to help people with disabilities or to run on different kinds of robots.
How it changes the race
This is a benchmark paper, not a new model. If the results hold, it gives developers of embodied SI a way to measure user-specific navigation, and it shows that general vision-language models do not yet handle it well.
Safety and control
The benchmark measures physical safety and collision risk. It also checks fit with user-specific constraints, which matters when SI guides people with impairments. The reported inconsistency across user profiles is a warning against relying on current models for this.
The abstract, by the authors
Accessible mobility is user-conditioned rather than scene-intrinsic, as walkable paths for one embodiment are inaccessible to another. Yet, existing benchmarks evaluate traversability under a fixed and limited embodiment. We present Vision-Language Accessible Navigation (VLAN), containing 1,000 scenes, 20K navigation trajectories, and 160K keyframes across diverse lighting, environmental conditions, scene semantics, and path geometries. For each trajectory, VLAN provides embodiment-specific actions and accessibility features for legged and wheeled agents, including humanoid and quadruped robots and people with visual or mobility impairments, while everyday obstacles and accessibility labels are categorized as Passable, Bypass, HighRisk, and Blocked. Unlike goal-only evaluation, our benchmark measures physical safety and alignment with user-specific constraints via goal-reaching, collision-aware, and accessibility-aligned metrics. We benchmark modern vision-language models and find they often fail to consistently adapt obstacle relevance, path selection, and safety criteria across user profiles, suggesting accessibility-aware navigation requires dedicated datasets, tasks, and metrics beyond conventional embodied navigation evaluation. All data and code open-source at https://github.com/InSAI-Lab/VLAN.
Quoted from arXiv:2611.18346. The research and its words belong to its authors.
