InterMASH: A Unified Geometric Representation for Grasp Synthesis

arXiv cs.GR details InterMASH, a new geometric representation for grasp synthesis that aims to bridge human and robotic hands with one compact encoding. The paper frames grasp generation as a hand-object-contact problem and replaces heavier, less complete descriptors with sphere-fixed anchors plus low-degree spherical harmonics.

That matters because grasp synthesis is only useful if the generated pose is both stable and physically plausible. InterMASH turns each anchor into an explicit token sequence, which makes the representation easier to interpret and better suited to transformer-style generation than dense implicit fields or contact maps alone.

The authors pair the representation with a conditional Diffusion Transformer that predicts hand geometry and contact jointly in the same space. On a large-scale ShadowHand benchmark, the method is reported to be competitive on key physical feasibility metrics, and it also supports joint training across multiple hands.

For developers working on robotics, simulation, or hand animation, the interesting bit is the cross-embodiment angle: fine-tuning with human grasp data reportedly improves robotic grasp success and diversity. If that holds up beyond the benchmark, it points toward a more reusable grasp pipeline for different hand morphologies instead of one-off models per rig.