A new research benchmark called VISTA is designed to evaluate AI systems that assist blind and low-vision users with everyday activities. The benchmark uses multimodal egocentric data—meaning first-person video and other sensor streams—to test goal-oriented perception.

The paper notes that while AI assistants could offer useful guidance, current benchmarks do not adequately test whether vision models can provide effective help in these scenarios. VISTA aims to fill that gap by focusing on tasks where users need real-time, context-aware support.

By grounding evaluation in egocentric, goal-oriented settings, VISTA could push vision-language models toward more practical assistance. The work is described in a preprint on arXiv; details of the dataset and evaluation protocol are not fully covered in the abstract.