Ranking-Based Robotics Benchmark (RB2)

RB2是一个专为机器人操作研究设计的基准数据集，由卡内基梅隆大学、纽约大学等机构合作创建。该数据集包含四个日常操作任务：倾倒、舀取、拉链和插入，旨在评估机器人在执行这些任务时的性能。数据集通过收集专家演示数据来训练机器人策略，并通过不同的测试场景评估这些策略的控制性能。RB2不仅提供了一套标准化的实验协议和最先进的机器人学习方法实现，还允许研究者在全球范围内比较和验证他们的算法，从而推动机器人操作研究的进步。

RB2 is a benchmark dataset dedicated to robotic manipulation research, co-developed by institutions including Carnegie Mellon University and New York University. This dataset encompasses four daily manipulation tasks: pouring, scooping, zipping, and insertion, with the goal of evaluating robotic performance when executing these tasks. It collects expert demonstration data to train robotic policies, and assesses the control performance of these policies across diverse test scenarios. Beyond providing a standardized experimental protocol and implementations of state-of-the-art robotic learning methods, RB2 also enables researchers worldwide to compare and validate their algorithms, thereby advancing robotic manipulation research.