Lingjing: New Multi-Agent Urban Simulation Testbed for Embodied AI

Developing intelligent systems that can coordinate various types of robots and autonomous vehicles in complex urban settings is a significant challenge. Existing simulation platforms often compartmentalize different robot types and decouple them from task design, limiting comprehensive evaluation. A new platform called Lingjing addresses these limitations by providing a unified simulation testbed for heterogeneous multi-agent embodied intelligence within open-ended urban environments. It achieves this by reconstructing evolving cities from real geographic data, synchronizing multiple physics engines, and exposing shared physical and structured urban state to the agents. Lingjing features a Gym-like interface, allowing users to define ReAct agents and create single or multi-agent missions using natural language, complete with configurable communication and resource constraints. Each simulation episode generates a detailed replay for systematic diagnosis, linking agent actions and communication to environmental changes and performance metrics. Initial evaluations using Lingjing highlight persistent issues in grounding and long-horizon execution for vision-language models, and reveal complex trade-offs in coordination and scalability.