DualWorldBench: Can Agents Plan Deliveries across Symbolic and Grounded Worlds?

DualWorldBench evaluates agent planning through three specialized modules: TextWorld for symbolic task composition, SimWorld for grounded topological path planning, and DualWorld-VQA for multimodal diagnostic testing. The benchmark tests an agent's ability to integrate perception and reasoning under varying temporal and physical constraints across symbolic and grounded environments.