Huajian Zeng¹, Lingyun Chen², Jiaqi Yang¹, Yuantai Zhang¹, Fan Shi³, Peidong Liu⁴, Xingxing Zuo^1,*
1Mohamed Bin Zayed University of Artificial Intelligence (MBZUAI), ²Technical University of Munich (TUM), ³National University of Singapore (NUS), ⁴Westlake University
^*Corresponding author

In Submission

FlowHOI is a two-stage flow-matching framework that generates semantically grounded, temporally coherent hand-object interaction (HOI) sequences for dexterous robot manipulation.

By decoupling geometry-centric grasping from semantics-centric manipulation and conditioning on 3D Gaussian splatting scene reconstruction, FlowHOI achieves 1.7× higher physics simulation success rate and 40× inference speedup compared to the strongest diffusion-based baseline.

These capabilities enable real-robot execution on diverse dexterous manipulation tasks, bridging the gap between HOI generation and practical robotic deployment.