Brief

Researchers have developed EquiDexFlow, a novel SE(3)-equivariant flow-matching model designed to generate dexterous grasps for robotic hands. Unlike previous methods that treat contact forces as a secondary verification step, EquiDexFlow integrates the prediction of wrist pose, joint angles, fingertip contacts, surface normals, and contact forces directly into the model. This approach ensures that predicted grasps are not only kinematically plausible but also physically stable, adhering to contact force conditions and friction constraints by construction. The model has demonstrated strong performance in simulations and on physical hardware, successfully enabling open-loop pick-and-hold trials. AI