VIMA-Bench

该数据集是一个模拟基准，包含了数千个通过程序生成的桌面任务，这些任务具有多模态提示，以及超过60万条专家轨迹用于模仿学习。此外，该基准还包含了一个四级评估协议，用于系统泛化，旨在测试模型的可扩展性和数据效率。该数据集规模宏大，包含数千个任务和超过60万条专家轨迹，其任务内容主要涉及使用多模态提示的机器人操作任务。This dataset is a simulated benchmark containing thousands of programmatically generated desktop tasks with multimodal prompts, as well as over 600,000 expert trajectories for imitation learning. Additionally, this benchmark includes a four-level evaluation protocol for systematic generalization, which aims to evaluate model scalability and data efficiency. As a large-scale dataset, it encompasses thousands of tasks and over 600,000 expert trajectories, with its core task domain focusing on robotic manipulation tasks leveraging multimodal prompts.