GOAG: Generative and Object-Agnostic Grasp Planner for Dexterous Robotic Manipulation
GOAG plans dexterous grasps without training on object-specific grasp data.
The paper frames grasping around shared surface geometry at gripper-object contact points. Its generative model learns a compact representation of a gripper’s contact surface distribution, then brings in object features only at inference. The authors report validation in simulation and real-world scenarios across different grippers, with an 86.93% average success rate on MultiDex objects. They say it generates many grasps faster while matching leading methods trained on that dataset. HF Daily Papers' note
The paper frames grasping around shared surface geometry at gripper-object contact points. Its generative model learns a compact representation of a gripper’s contact surface distribution, then brings in object features only at inference. The authors report validation in simulation and real-world scenarios across different grippers, with an 86.93% average success rate on MultiDex objects. They say it generates many grasps faster while matching leading methods trained on that dataset. HF Daily Papers' note
score 4