CARG-grasp: constraint-aware region-guided multi-scale 6-DoF grasping for industrial parts
Purpose Industrial parts often involve multi-scale geometries, cluttered stacking, severe occlusions, fragmented local features and complex operational constraints, which make safe six-degree-of-freedom (6-DoF) grasping difficult. This paper aims to propose CARG-Grasp, a constraint-aware region-guided multi-scale 6-DoF grasping method for complex industrial parts. Design/methodology/approach The method consists of a constraint-aware feasible region generation module (CFRG) and a multi-scale self-attention heatmap-guided grasp detection network (MSA-HGGD). CFRG converts coarse task-constrained regions into feasible grasping regions with geometric consistency and physical safety through boundary refinement, multi-constraint fusion and depth-adaptive safety field modelling. MSA-HGGD grasp detection module then predicts 6-DoF grasp poses within the feasible regions. A three-branch multi-scale feature extraction module (TBMSE) and a cross-scale adaptive attention fusion module (CSAAF) are used to jointly model fine local structures, medium-scale contour features and global spatial relationships. Findings Experimental results demonstrate that CARG-Grasp achieves competitive grasp detection performance while significantly improving grasp feasibility, constraint satisfaction and real-world execution robustness compared with comparative methods. The results demonstrate enhanced grasping safety and reliability in cluttered industrial-part manipulation scenarios. Originality/value This study integrates feasible region generation, depth-adaptive safety modelling and multi-scale heatmap-guided grasp detection into a unified 6-DoF grasping framework. The proposed method explicitly incorporates task constraints and physical safety into grasp generation, making it suitable for practical industrial robotic applications.
Authors
- Shuangshuang Wu (ORCID: https://orcid.org/0000-0002-4226-2039)
- Huaidong Zhou (ORCID: https://orcid.org/0000-0002-7106-6555)
- Wenbai Chen (ORCID: https://orcid.org/0000-0001-7683-2776)
- Xintong Meng (ORCID: https://orcid.org/0009-0001-9003-8613)
- Hang Ma (ORCID: https://orcid.org/0009-0003-8681-5269)
- Peiyuan Zhao (ORCID: https://orcid.org/0009-0008-0740-299X)
Institutions
- Beijing Information Science & Technology University (CN)
- Tsinghua University (CN)
Publication Details
- Journal
- Industrial Robot the international journal of robotics research and application
- Published
- 2026-11-19
- DOI
- https://doi.org/10.1108/ir-07-2026-0354
- Primary Topic
- Robot Manipulation and Learning
- Type
- article
- Field-Weighted Citation Impact
- 0.00