CongLab-Research/LabHorizon-3D-Asset-Perception
收藏资源简介:
该数据集是LabHorizon项目的Level 1部分,专注于实验室3D资产感知和协议条件下的下一个动作预测。每个样本包含同一实验室资产的三个渲染视图、历史实验动作以及一组候选下一个动作。任务目标是选择与实验协议一致的下一个动作。该任务并非通用的图像描述或视觉问答,而是要求模型将实验室3D感知与协议条件下的动作预测相结合:视觉资产需与实验历史匹配,所选下一个动作需与邻近协议步骤和细粒度参数保持一致。数据集包含3,000个训练样本和200个测试样本,输入包括资产图像、历史动作和候选下一个动作,输出为正确的下一个动作。
This dataset is the Level 1 split of LabHorizon, focusing on laboratory 3D asset perception and protocol-conditioned next-action prediction. Each example pairs three rendered views of the same laboratory asset with historical experimental actions and a set of candidate next actions. The target is the protocol-consistent next action. The task is not generic image captioning or visual question answering; it asks whether a model can connect laboratory 3D perception with protocol-conditioned action prediction: the visual asset should match the experimental history, and the selected next action should be consistent with nearby protocol steps and fine-grained parameters. The dataset includes 3,000 training samples and 200 test samples, with inputs comprising asset images, historical actions, and candidate next actions, and the output being the gold next action.




