arXivDaily arXiv每日学术速递 周一至周五更新
arXiv周末暂无论文更新,休息一下吧,周末愉快~~
arXiv 2610.26962cs.RO
收藏

DexGraspFlow:用于跨 embodiment 灵巧抓取生成的拓扑感知 SE(3) 流匹配

DexGraspFlow: Topology-Aware SE(3) Flow Matching for Cross-Embodiment Dexterous Grasp Generation

Nanyang Technological University(南洋理工大学);ROKAE Robotics(珞石机器人)

Bin Zhou, Boyu Ma, Xinyu Zhou, Jiahua Dong, Chuhao Zhou, Yizhou Liu, Hao Wan, Jianfei Yang

AI总结:

提出 DexGraspFlow 框架,结合几何与运动学拓扑实现跨 embodiment 灵巧抓取,仿真成功率 95.33%,延迟降 49%,真实机器人试验成功率 91%,支持零样本适配与可控生成。

摘要:

手部几何结构、关节参数化及运动学拓扑的差异给跨 embodiment 灵巧抓取生成带来挑战,仅连杆间的空间关系不足以完整捕捉手部运动学拓扑。本文提出 DexGraspFlow,一种结合几何关系与运动学拓扑的条件式连杆级 SE(3) 流匹配框架,该框架编码局部物体几何、连杆几何及当前位姿;双流 DiT 模型连杆-连杆协调与连杆-物体交互,源自运动树距离的注意力偏置为建模演化几何交互提供显式结构信息。模型联合学习平移与旋转速度场,将其整合生成连杆位姿,通过逆运动学恢复关节配置;推理时的手掌方向引导可实现可控抓取生成,无需重新训练。在三种机器人手部及 30 组物体-手部对测试中,DexGraspFlow 平均仿真成功率达 95.33%,含逆运动学的平均生成延迟比现有最优基线低约 49%;进一步实验表明其对七种手指移除变体具备零样本适应性,且支持手掌方向控制;针对部署手部的演示训练模型,在 100 次真实机器人抓取试验中成功率达 91%,验证了该框架在物理抓取执行中的适用性。

英文摘要:

Differences in hand geometry, joint parameterization, and kinematic topology pose challenges for cross-embodiment dexterous grasp generation. However, spatial relations among links alone are insufficient to fully capture the hand's kinematic topology. We propose DexGraspFlow, a conditional link-level $\mathrm{SE}(3)$ flow-matching framework integrating geometric relations and kinematic topology. The framework encodes local object geometry, link geometry, and current poses. A dual-stream DiT models link--link coordination and link--object interaction, while attention biases derived from kinematic-tree distances provide explicit structural information for modeling evolving geometric interactions. The model jointly learns translational and rotational velocity fields, integrates them to generate link poses, and recovers joint configurations through inverse kinematics. Additionally, inference-time palm-orientation guidance enables controllable grasp generation without retraining. Across three robotic hands and 30 object--hand pairs, DexGraspFlow achieves a 95.33\% average simulation success rate, with an average generation latency, including inverse kinematics, approximately 49\% lower than that of a state-of-the-art baseline. Further experiments demonstrate zero-shot adaptability to seven finger-removal variants and palm-orientation control. A model trained on demonstrations for the deployed hand achieves a 91\% success rate over 100 real-robot grasping trials, supporting the framework's applicability to physical grasp execution.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑