Huiqiong0124/RoboTrustBench_Dataset
收藏资源简介:
RoboTrustBench_Dataset是RoboTrustBench的一个子数据集,RoboTrustBench是一个用于评估视觉语言模型是否能在正常、约束敏感、反事实和对抗性设置下可靠理解机器人操作指令和视觉场景的基准。本次发布的不是完整数据集,它包含一个40个样本的子集,完整的RoboTrustBench数据集将很快发布。每个样本将初始图像与语言指令配对,覆盖四个以可信度为导向的场景类别:正常场景(普通机器人操作指令)、约束敏感场景(需要关注模糊性、遮挡、障碍、干扰物或轨迹约束的指令)、反事实场景(与视觉或物理可行性冲突的指令)和对抗性场景(涉及人类攻击或环境损害的不安全或有害指令)。数据集当前包含40个样本,组织成四个主要类别,包括正常(10个样本)、约束敏感(10个样本,细分为多个子类别)、反事实(10个样本,细分为多个子类别)和对抗性(10个样本,细分为攻击人类和环境损害)。数据集结构与Hugging Face的imagefolder加载器兼容,metadata.jsonl文件包含指向data/目录下每个图像的file_name字段。
RoboTrustBench_Dataset is a subdataset of RoboTrustBench, a benchmark for assessing whether vision-language models can reliably understand robotic manipulation instructions and visual scenes across normal, constraint-sensitive, counterfactual, and adversarial settings. This release is not the full dataset. It contains a 40-sample subset, and the complete RoboTrustBench dataset will be released soon. Each sample pairs a initial image with a language instruction, covering four trustworthiness-oriented scenario categories: Normal scenarios (ordinary robotic manipulation instructions), Constraint-sensitive scenarios (instructions requiring attention to ambiguity, occlusion, obstacles, distractors, or trajectory constraints), Counterfactual scenarios (instructions that conflict with visual or physical feasibility), and Adversarial scenarios (unsafe or harmful instructions involving human attack or environmental damage). The current dataset contains 40 samples organized into four major categories, including Normal (10 samples), Constraint-sensitive (10 samples with multiple subcategories), Counterfactual (10 samples with multiple subcategories), and Adversarial (10 samples subdivided into attack human and environmental damage). The dataset structure is compatible with the Hugging Face imagefolder loader because metadata.jsonl contains a file_name field pointing to each image under data/.





