Kinetix Benchmark

该数据集是一个包含12个高度动态和随机环境的基准，旨在评估在模拟推理延迟条件下实时分块（RTC）算法的性能。这些环境涵盖了诸如投掷、接球和平衡等动态动作，并添加了高斯噪声以模拟不完美的执行效果。该数据集由6个专家策略在12个环境中生成的100万个转换构成，任务是评估在动态环境中的实时动作分块效果。

This dataset is a benchmark encompassing 12 highly dynamic and stochastic environments, which is designed to evaluate the performance of real-time chunking (RTC) algorithms under simulated inference latency conditions. These environments cover dynamic tasks including throwing, catching, and balancing, with Gaussian noise added to simulate imperfect execution outcomes. The dataset comprises 1 million transitions generated by 6 expert policies across the 12 environments, with its core task being to assess the effectiveness of real-time action chunking in dynamic scenarios.