Event

Proposes cross-embodiment dexterous grasp generation framework DexGraspFlow

1 reports · 1 sources

Bin Zhou et al. propose DexGraspFlow, a conditioned link-level SE(3) flow matching framework for cross-embodiment dexterous grasp generation, which combines geometric relations with kinematic topology, uses a dual-stream DiT to model inter-link coordination and link-object interaction, and provides explicit structural information via attention biases derived from kinematic tree distances. On three robotic hands and 30 object-hand pairs, DexGraspFlow achieves an average simulation success rate of 95.33%, and its average generation latency, including inverse kinematics, is about 49% lower than a state-of-the-art baseline. The experiments also demonstrate its zero-shot adaptability to seven finger-removal variants and palm orientation control. A model trained with deployed hand demonstration data achieved a 91% success rate in 100 real-robot grasp trials.