Grasp Pose Refinement via Object-centric Scoring

Accurate grasp pose prediction is critical for reliable robotic manipulation, particularly in heterogeneous, low-volume manufacturing environments. While GraspNet enables scene-level grasp prediction based on geometry, it lacks object-centered grasp pose refinement framework that integrates 6D pose estimation, GraspNet-Baseline, and a lightweight grasp pose refinement module. The target object’s 6D pose is estimated, and its CAD model is aligned to the scene, allowing the generation of grasp candidates, even in occluded or out-of-view regions. These grasp candidates, originally generated by GraspNet in the scene frame, are transformed into the object frame and re-score using our refinement module. This module is trained, to approximate a rule-based, object-centric grasp scoring function that accounts for alignment, distance to the object’s center, and approach direction. The module is implemented as a lightweight multilayer perceptron(MLP) allowing fast, generalized refinement without retraining GraspNet. Our framework bridges scene-level grasp generation with object-level grasp reasoning, enhancing grasp stability in occluded and complex environments.