RoboCerebra

RoboCerebra是一个大规模的机器人操作评估数据集，旨在评价长期 horizon 的机器人操作能力。数据集由北京航空航天大学、新加坡国立大学和上海交通大学的研究团队创建，包含1000个条目，可在robocerebra.github.io访问。数据集利用GPT模型生成高水平的任务指令，并将其分解为子任务序列，通过人类操作者在模拟环境中执行子任务，收集轨迹数据。数据集的特点是轨迹长度显著延长（约6倍于现有数据集），并具有丰富的子任务注释，为评估长期 horizon 的推理能力提供了一个更加严格和全面的测试平台。

RoboCerebra is a large-scale robotic manipulation evaluation dataset designed to assess long-horizon robotic manipulation capabilities. The dataset was created by research teams from Beihang University, the National University of Singapore, and Shanghai Jiao Tong University, contains 1000 entries, and is accessible at robocerebra.github.io. The dataset leverages GPT models to generate high-quality task instructions and decompose them into subtask sequences. Trajectory data is collected by having human operators perform the subtasks in simulated environments. The dataset features significantly extended trajectory lengths (approximately 6 times that of existing datasets) and rich subtask annotations, serving as a more rigorous and comprehensive testbed for evaluating long-horizon reasoning capabilities.