InternRobotics/IROS-2025-Challenge-Manip

IROS-2025-Challenge-Manip数据集是针对机器人操作挑战的一个基准数据集，包含预训练、训练和验证三个部分。预训练集包含大约20000个单次抓取和放置轨迹；训练集为每个任务提供了大约100个轨迹；验证集包括测试时间的场景和对象资源，以USD格式存储。每个轨迹都包含多视角视频记录、机器人状态和对应任务执行的动作。该数据集旨在支持在IROS挑战环境下的机器人操作预训练、任务特定微调和评估。The IROS-2025-Challenge-Manip dataset is a benchmark for robotic manipulation challenges, consisting of pretrain, train, and validation splits. The pretrain split contains approximately 20,000 single pick-and-place trajectories; the train split provides about 100 trajectories per task; and the validation split includes test-time scenes and object assets in USD format. Each trajectory comprises multi-view video recordings, robot states, and actions corresponding to task execution. This dataset is designed to support pretraining, task-specific fine-tuning, and evaluation for robotic manipulation in the IROS Challenge setting.