Contact2Motion: Contact guided dexterous grasp motion generation with synergy embedded optimization

Abstract

Existing work on grasp generation has shown promising results for parallel or three-finger grippers by combining grasp pose synthesis and motion planning. However, applying this framework to anthropomorphic multi-fingered hands remains challenging due to their high degrees of freedom (DoFs) and complex contact patterns. The large DoF results in a high-dimensional search space, and grasp stability is sensitive to subtle variations in contact. In this paper, we propose a two-stage pipeline for dexterous grasp motion generation that integrates contact reasoning with spatial and temporal hand synergy. In the pose synthesis stage, we factorize the object-to-grasp generation into object-to-contact map prediction and contact-to-pose mapping, reducing the complexity of direct grasp pose generation. A novel diffusion- based contact generator is trained on a low-dimensional latent manifold to improve sample quality and diversity. In the motion planning stage, we introduce an optimization framework with synergy-aware parameterization. Spatial synergy is encoded via a neural latent space, while temporal synergy is represented by a compact parametric model controlling motion timing. This reduces optimization complexity and avoids unnatural motions. Additionally, we propose a contact- guided local refinement module and a joint optimization strategy to mitigate errors from the staged pipeline. Extensive experiments demonstrate that our method achieves significantly better performance than existing 4D grasp motion baselines, improving success rate by 40% and reducing planning time from 30 seconds to under 2 seconds. Using our pipeline, we further construct a large-scale grasp motion dataset with over one million sequences across 8000 objects and multiple hand types.