DEAL-Grasp: Decoupled Alignment Representation for Geometry-Aware Dexterous Grasp Generation

arXiv cs.GR details DEAL-Grasp, a new approach to dexterous grasp synthesis aimed at more believable hand-object interaction for VR, embodied intelligence, and digital human pipelines. The core idea is to decouple global rigid motion from local hand articulation, then recover the rigid transform through closed-form Procrustes alignment instead of regressing everything in one coupled pose space.

That separation matters because coupled pose generation tends to produce unstable samples and awkward contacts, especially when the hand has many degrees of freedom. DEAL-Grasp models the mixed state with heterogeneous-state flow matching and component-wise vector fields, while adding time-adaptive physical regularization during training to keep contacts plausible.

At inference, the system generates grasps by integrating the learned vector field directly, avoiding test-time optimization and extra physical guidance. The paper reports strong force-perturbation success, low penetration, and high diversity on MultiDex and zero-shot RealDex, along with lower latency than optimization-heavy baselines.

For game teams, the practical angle is clear: better hand-object interaction data and faster synthesis could help with VR avatars, digital humans, robotics-inspired animation, and any pipeline that needs convincing grasp poses without expensive per-sample cleanup.