Maniwhere

该数据集提供了一个通用的视觉强化学习框架，使得训练出的机器人策略能够在多种开放世界场景中，跨越不同类型的视觉干扰实现泛化。此外，该数据集支持仿真到现实的迁移能力，并允许在不同视觉外观和机体上进行评估。实验范围涵盖了8个任务和3种硬件平台。这些任务包括机器人操作，如关节物体的操作、双手操作以及灵巧手操作等。

This dataset presents a general visual reinforcement learning framework, enabling the trained robotic policies to generalize across diverse visual disturbances in multiple open-world scenarios. Furthermore, this dataset supports sim-to-real transfer capabilities and allows evaluation across different visual appearances and robot morphologies. The experimental scope covers 8 tasks and 3 hardware platforms. These tasks include robotic manipulation, such as manipulation of articulated objects, dual-arm manipulation, and dexterous hand manipulation, among others.