sarahnator/genesis-physical-interventions

该数据集包含受控的Genesis模拟，用于在具身AI环境中评估物理可行的世界模型。具体场景包括Ramp-cup-water（斜坡-杯子-水）、Robotic pour（机器人倒水）和Pendulum（钟摆）。任务类型涵盖单次推出结果预测、标量物理预测、逆参数预测和成对反事实比较。数据集包含111个VLM条目，按场景和查询类型分布，文件结构包括数据文件、场景和人工制品（如帧、视频和轨迹）。这是一个研究基准，用于物理推理，而非完整的真实世界机器人基准，标签基于模拟器而非真实测量，部分标签使用代理（例如机器人倒水中的接收器分数）。This dataset contains controlled Genesis simulations for evaluating physically viable world models in embodied AI settings. Scenes include Ramp-cup-water, Robotic pour, and Pendulum. Task types include single-rollout outcome prediction, scalar physical prediction, inverse parameter prediction, and pairwise counterfactual comparison. It consists of 111 VLM entries distributed across scenes and query types, with files such as data files, scenes, and artifacts (e.g., frames, videos, trajectories). It is a research benchmark for physical reasoning, not a complete real-world robotics benchmark, with simulator-derived labels and some using proxies like receiver_fraction for robotic pouring.