robovlms/bytedance_robot_benchmark_20

ByteDance机器人基准数据集（BDRBench-20）是一个包含8K高质量轨迹的视觉-语言-动作（VLA）数据集，旨在评估VLA模型在现实世界场景中的性能。该数据集包括20个常见的操作任务，如取放、倒水和开关动作。数据集分为训练集和验证集，每个子目录下包含注释文件和视频文件。注释文件中包含任务的英文描述、视频路径、动作和状态等信息。视频文件包括静态相机录制的视频和手腕相机录制的视频。数据集还提供了每个任务的视频轨迹和注释文件。

ByteDance Robot Benchmark (BDRBench-20) is a vision-language-action (VLA) dataset containing 8K high-quality trajectories. It was created to evaluate the performance of VLA models in real-world scenarios. This dataset includes 20 common manipulation tasks, such as pick-and-place, pouring, and open/close actions. The dataset is divided into `train` and `val` sets, with two subdirectories: `anns` (annotations) and `media` (videos). The `anns` directory contains annotation files for each subtask, while the `media` directory contains rollout videos for each task.