未明确提供

本论文提出了一种新的强化学习（RL）环境，旨在通过引入自然世界的复杂性来改进RL算法的鲁棒性。该环境包含三个系列的任务，主要集中在视觉推理任务上，如图像分类和对象定位，以及对现有RL基准任务的修改，引入自然视频作为背景。这些任务的设计旨在支持快速数据采集，实现公平的训练/测试分离，并便于结果的比较和复制。通过这种方式，论文鼓励RL研究社区开发更稳健的算法，以满足高标准的评估要求。

This paper proposes a novel reinforcement learning (RL) environment aimed at enhancing the robustness of RL algorithms by incorporating the complexity of the natural world. The environment includes three series of tasks, primarily focusing on visual reasoning tasks such as image classification and object localization, alongside modifications to existing RL benchmark tasks that adopt natural videos as backgrounds. These tasks are designed to support rapid data collection, enable fair train/test splits, and facilitate result comparison and reproducibility. In this manner, the paper encourages the RL research community to develop more robust algorithms that meet high-standard evaluation requirements.