RG-DGTR: Robust Geometric-Enhanced Dexterous Grasp Transformer with Robust Adaptation

Abstract

Synthesizing dexterous grasp poses from 3D point clouds remains a central challenge in robotic manipulation. While Transformer-based set prediction paradigms have significantly enhanced inference efficiency, they remain constrained by primary robustness bottlenecks within the complex mapping of high-dimensional configuration space. Specifically, local feature aggregation tends to over-smooth high-frequency curvature priors, while limited receptive fields often result in a lack of global constraints for multi-finger kinematic coordination. Furthermore, existing Test-Time Adaptation (TTA) mechanisms exhibit high sensitivity to non-ideal initial poses, which frequently leads to distortion. To address these limitations, we propose RG-DGTR, a robust geometric-enhanced framework for dexterous grasp generation. Our framework first incorporates a geometric curvature prior encoder that utilizes explicit second-order curvature as a physical bias to guide grasp queries toward regions with high force-closure potential. We then introduce a spatial global correlation module that leverages long-range interactions to overcome the deficiencies of local operators in establishing global associations. Finally, we present a robust anchor-based adaptation strategy with failure-sensing capability, which constructs a dynamic gravity-traction trust region to effectively suppress optimization divergence under poor initializations while balancing global topology maintenance with local contact refinement. Extensive experiments on the large-scale DexGraspNet benchmark demonstrate that our approach significantly improves both grasp quality and physical consistency.