UnrealZoo is a dataset of over 100 realistic 3D virtual worlds built with Unreal Engine. It provides a diverse set of environments and interactive objects, including humans, animals, robots, and vehicles, making it ideal for embedded AI research. It extends UnrealCV to provide optimized APIs and tools for data collection, environment augmentation, distributed training, and benchmarking, while significantly improving rendering and communication efficiency to support advanced applications such as multi-agent interaction. Experimental evaluations on visual exploration and tracking tasks reveal that while environmental diversity offers significant benefits for developing generalizable reinforcement learning agents, current embedded agents face persistent challenges in open-world scenarios, such as navigating unstructured terrain, adapting to unseen shapes, and managing latency in closed-loop control systems when interacting with dynamic objects. Therefore, UnrealZoo serves as a comprehensive testing environment and pathway for real-world deployment of more capable embedded AI systems.
