d3d3shan/pi05-libero-plus-robot-initial-failures
收藏资源简介:
该数据集名为pi0.5 LIBERO-plus机器人初始状态失败,是一个用于评估TensorAuto/tPi0.5-libero策略在LIBERO-plus基准测试的机器人初始状态扰动子集上的失败摘要数据集。它包含失败视频网格、元数据文件和日志,用于分析机器人初始状态导致的失败情况。评估设置包括使用libero_10套件、393个任务、每个任务5个情节、总2430个评估情节,情节长度为520步,硬件为RTX 5070 Ti笔记本电脑GPU。数据统计显示成功率为71.52%,失败692次,其中504个失败视频被渲染成网格,71个失败缺少视频。数据集文件包括视频目录、机器可读的失败记录JSON文件、人类可读的图例MD文件、低难度失败任务重运行清单JSON、评估配置JSON和本地队列日志。
The dataset is named pi0.5 LIBERO-plus Robot Initial State Failures and is a collection of failure summary artifacts for evaluating the TensorAuto/tPi0.5-libero policy on the LIBERO-plus benchmarks Robot Initial States perturbation subset. It includes failure grid videos, metadata files, and logs for analyzing failures caused by robot initial states. The evaluation setup uses the libero_10 suite, 393 tasks, 5 episodes per task, totaling 2430 evaluated episodes, with an episode length of 520 steps, and hardware being an RTX 5070 Ti laptop GPU. Statistics show a success rate of 71.52%, with 692 failures, of which 504 failure videos are rendered into grids, and 71 failures are missing videos. Dataset files include a videos directory, machine-readable failure records JSON, human-readable legend MD, low-difficulty failed-task rerun manifest JSON, evaluation configuration JSON, and a local queue log.




