A new arXiv preprint presents a Persian vision-language-action dataset and benchmark for grounding natural-language instructions into executable action plans conditioned on scene images. This kind of grounding is a core capability for embodied human-robot interaction, but according to the abstract, existing benchmarks largely target English, simulated, or video-grounded environments.

The new benchmark aims to fill that gap by evaluating models in Persian and by paying special attention to ambiguous and infeasible instructions. That design matters for real-world robotics, where user commands are often unclear, underspecified, or impossible to execute in the observed scene.

The source is the paper's abstract, so no dataset size, model results, or evaluation details are yet available from this material. The contribution is presented as a benchmark and dataset rather than as a proposed method.