Goal Force synthetic causal primitives dataset
收藏资源简介:
Goal Force合成因果基元数据集由布朗大学与康奈尔大学联合创建,包含12,000条物理交互视频数据,涵盖弹性碰撞、多米诺骨牌效应等基础物理现象。数据集包含三个子集:3,000条多米诺骨牌链式反应视频、6,000个滚动球体碰撞案例以及3,000朵康乃馨的非刚性动力学模拟,所有数据均通过Blender和PhysDreamer物理引擎生成。该数据集通过多通道物理控制信号(直接力/目标力/质量)标注,专门用于训练视频生成模型理解物理因果关系。其核心应用在于构建神经物理模拟器,实现无需外部引擎的物理感知规划,可扩展至机器人操作、工具使用等复杂场景的零样本迁移。
The Goal Force Synthetic Causal Primitives Dataset, jointly created by Brown University and Cornell University, comprises 12,000 physically interactive video clips covering fundamental physical phenomena including elastic collisions and domino effects. The dataset contains three subsets: 3,000 videos of domino chain reactions, 6,000 rolling sphere collision cases, and 3,000 non-rigid dynamic simulations of carnations, with all data generated using the Blender and PhysDreamer physics engines. Annotated with multi-channel physical control signals (direct force, target force, and mass), this dataset is specifically designed to train video generation models to understand physical causal relationships. Its core applications lie in building neural physics simulators, enabling physics-aware planning without external engines, and supporting zero-shot transfer to complex scenarios such as robot manipulation and tool use.

- 1Goal Force: Teaching Video Models To Accomplish Physics-Conditioned Goals布朗大学; 康奈尔大学 · 2026年



