CASHER

该数据集名为CASHER，是一种针对机器人操作的大规模持续数据收集流程。它采用了一种多任务、多环境的实到仿真再到实的方法，利用真实世界场景的数字孪生在模拟中生成数据。该数据集使得在新环境中能够通过3D扫描和最少的人工演示高效地进行微调，并且其设计目标是可以扩展到成千上万的非专家和真实环境。该数据集的任务是机器人操作和学习可泛化的决策策略。

The dataset named CASHER is a large-scale continuous data collection pipeline designed for robotic manipulation. It adopts a multi-task, multi-environment real-to-sim-to-real approach, leveraging digital twins of real-world scenes to generate data within simulated environments. This dataset enables efficient fine-tuning in novel environments via 3D scanning and minimal human demonstrations, and it is designed to scale to thousands of non-experts and real-world scenarios. The core tasks of this dataset focus on robotic manipulation and learning generalizable decision-making policies.