Benchmark reveals challenges in multi-goal agent delivery planning
DualWorldBench: Can Agents Plan Deliveries across Symbolic and Grounded Worlds?
Artificial Intelligence
Summary
Planning for a robot or virtual agent to pick up and deliver items in a complex environment is hard because it requires understanding both abstract task details and real-world navigation. The authors created DualWorldBench, a set of tests that check how well agents can plan using symbolic data and real sensory information together. Their results show that agents often fail to make effective plans when tasks have many steps or when observations don’t easily translate to actions. This benchmark helps identify where current technology struggles and encourages better integration of perception and planning.
What this means in practice
- •For robotics software teams: Evaluate and improve robot navigation and task planning algorithms for delivery tasks that combine abstract goals and actual environment layout.
- •For autonomous vehicle developers: Test autonomous systems' ability to integrate sensory data with high-level routing and task schedules under complex constraints.
Authors
Yang Chen, Bin Wen, Danyang Peng, Hong-Jie You, Zi-Qiao Shang, Lan-Zhe Guo
Abstract
Recent multimodal agents have shown promising progress in embodied planning and web interaction. However, existing evaluations often rely on simplified tasks with limited combinatorial structure, leaving it unclear whether current agents can robustly integrate perception and reasoning for multi-goal planning under temporal, physical, and multimodal constraints. We introduce DualWorldBench, a unified benchmark for evaluating agent planning under controlled symbolic--temporal and grounded--topological complexity. Built around a shared pickup--delivery interface, DualWorldBench consists of three complementary components: TextWorld, which isolates symbolic planning under varying task compositions and release dynamics; SimWorld, which evaluates grounded path planning under controlled topological variation with text, image, or multimodal inputs; and DualWorld-VQA, a diagnostic suite spanning perception, scene understanding, and planning. Experiments reveal a common gap between local understanding and globally consistent planning. In TextWorld, agents struggle to derive optimal task sequences under interacting temporal constraints; in SimWorld, grounded observations are not reliably converted into feasible action sequences. DualWorld-VQA further confirms that perception and topology understanding do not necessarily translate into constrained planning. Together, these results identify the central challenge as composing symbolic or grounded evidence into globally feasible and efficient multi-goal plans. These findings establish DualWorldBench as a challenging testbed for agents with tighter perception--reasoning--planning integration. All datasets and code are publicly available at our project website https://dualworld.netlify.app/.