A2A Benchmark for Vision-and-Language Navigation in Agricultural Robotics
October 22, 2026
The A2A benchmark introduces 1,560 navigation episodes across six agricultural scene classes using continuous actions and step-by-step instructions. Evaluation shows current state-of-the-art VLN methods fail on long-form instructions, necessitating new modules like STL to improve performance in complex farming environments.
HOW THIS AFFECTS YOU
●
builderYou can use this benchmark to test robot navigation reliability in unstructured agricultural settings.
●
researcherThis provides a specialized dataset to address the performance gap in long-form instruction following for VLN.