赋予智能体通过与环境交互自主学习的能力是实现下一代人工智能的关键。本文，我们介绍了一种基于虚幻4的虚拟仿真环境，用于训练和测试自主智能体。该环境具有高逼真、可交互、灵活通用的特点，使得智能体能够在其中自由探索，自主学习场景感知、常识推理、决策控制等多项能力。为了验证该环境的可用性，我们用实验演示了如何在虚拟环境中构建自主智能，即利用强化学习方法训练端到端的神经网络实现基于视觉感知的目标搜索和目标追踪任务。

The key to realizing the next generation of artificial intelligence is how to make agents to autonomous learning from the interaction with the environment. In this paper, we introduce a collection of UE4 (Unreal Engine 4)-based virtual environments for training and testing the autonomous agents. The environment is high-fidelity, interactable, and flexible to meet the requirements of enabling agents to freely explore and autonomously learn the ability of scene perception, common sense, and decision making. To validate the usability of the environments, we developed agents via reinforcement learning to perform object searching and active object tracking.