BiDexGrasp: Coordinated Bimanual Dexterous Grasps across Object Geometries and Sizes

A research team led by Mu Lin, Yi-Lin Wei, and eight other collaborators has introduced BiDexGrasp, a breakthrough system addressing the critical bottleneck in bimanual robotic manipulation: the lack of comprehensive training data. Their novel two-stage synthesis pipeline first performs efficient region-based grasp initialization, then applies decoupled force-closure optimization to generate physically feasible grasps at scale. This enabled the creation of an unprecedented dataset containing 9.7 million annotated grasp configurations across 6,351 objects ranging from 30 to 80 centimeters in size.

Building on this dataset, the team developed a generative framework with two key innovations: a bimanual coordination module that ensures both hands work in concert, and a geometry-size-adaptive strategy that generalizes to unseen objects. The system demonstrated superior performance in extensive experiments, successfully transferring from simulation to real-world robotic platforms. This represents a significant advancement toward robots that can handle complex, everyday objects with human-like dexterity using both hands simultaneously.