Multi-Embodiment Tabletop Grasping Dataset (SeededGrasp dataset)
收藏资源简介:
该数据集是由多伦多大学与Vector Institute等机构联合创建的首个多末端执行器桌面抓取数据集,专门针对复杂杂乱场景中的机器人抓取任务设计。数据集规模达256万个抓取姿态,覆盖610个杂乱场景、334个物体对象以及Franka Panda、Robotiq 3-Finger和Allegro三种典型夹爪,数据通过合成生成流程从MultiGripperGrasp数据集转化而来。创建过程采用物理仿真与启发式碰撞检测相结合的方法,将稳定抓取姿态适配到桌面杂乱场景中,并生成2048点云作为场景表示。该数据集主要应用于多模态机器人抓取研究领域,旨在解决语言引导的多末端执行器抓取在复杂场景中的泛化问题,为分离语义推理与几何执行的框架提供训练基础。
This dataset is the first multi-end-effector desktop grasping dataset jointly created by the University of Toronto, Vector Institute and other institutions, specifically tailored for robotic grasping tasks in complex cluttered scenes. It contains 2.56 million grasping poses, covering 610 cluttered scenes, 334 distinct objects, and three representative grippers: Franka Panda, Robotiq 3-Finger and Allegro. This dataset is derived from the MultiGripperGrasp dataset through a synthetic generation pipeline. The development process employs a hybrid approach combining physics simulation and heuristic collision detection, which adapts stable grasping poses to cluttered desktop scenes and generates 2048-point point clouds as scene representations. This dataset is primarily utilized in the field of multimodal robotic grasping research, aiming to address the generalization challenge of language-guided multi-end-effector grasping in complex scenes, and provide training foundations for frameworks that separate semantic reasoning and geometric execution.

- 1SeededGrasp: Language-Guided Grasping in Complex Scenes with Multiple Embodiments多伦多大学; Vector Institute; 不列颠哥伦比亚大学; Google DeepMind · 2026年




