Answers are grounded in this paper and cite their sources.
Robotic grasping of irregular coal gangue objects is challenging because incomplete geometric observations and execution errors can compromise the reliability of initially predicted grasps. Sensor-based multimodal perception provides essential geometric and physical information for reliable robotic manipulation in complex coal gangue environments. This paper presents a geometric-prior-guided grasp refinement framework that couples bounded residual reinforcement learning with visual–physical representation alignment. GraspNet-generated six-degree-of-freedom grasp candidates provide nominal poses, around which a PPO-based policy predicts local translational and rotational corrections. Cross-modal contrastive learning aligns point-cloud features with the corresponding robot physical states to support residual-action prediction, while a joint policy-learning objective incorporates contrastive and action-regularization terms to discourage excessive corrections. The framework therefore connects geometric grasp initialization, state-conditioned pose refinement, and representation learning within a common optimization process. In the reported MuJoCo simulation, the full model achieves 98.0% grasp success versus 97.5% for Baseline PPO, with the reported convergence measure decreasing from 1200 to 1000 training rounds. In the reported CR5 laboratory evaluation, the full model achieves an 86% grasping success rate, compared with 70% for Baseline PPO, while terminal position deviation decreases from 0.046 m to 0.031 m. These results indicate that geometric-prior-guided residual refinement can improve robotic grasping performance under the evaluated laboratory conditions, particularly in terms of physical grasping success rate and terminal position accuracy.
No takes yet. Share an insight, caveat, or question.
Wang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: