Boundary-Aware Multi-Task Knowledge Distillation for Lightweight Grasp Detection
Lightweight grasp detectors suit resource-constrained robotic platforms, but compact encoders lose the spatial discrimination needed to keep grasp quality, orientation, and opening width consistent at one location, and generic distillation transfers teacher outputs without testing whether their grasp geometry is valid. This paper presents BA-MTKD, a boundary-aware multi-task distillation framework that transfers only teacher outputs whose quality, angle, and width agree with the training annotation and couples quality transitions with local angle–width relations at grasp boundaries. All teacher-dependent components are discarded after training. Against sixteen distillation methods retrained under one matched RGB protocol, a MobileNetV3-Small student trained with BA-MTKD achieves the highest accuracy on Cornell, OCID-Grasp, and VMRD V2 (97.6%, 84.7%, and 77.9%) and recovers 84–90% of the teacher–student gap. Its advantage widens under the strict criterion (IoU 0.50, 15°): it reaches 89.1%, 68.4%, and 59.8%, exceeds the strongest competitor by 2.0–3.3 points, and comes within 0.9 points of a twelve-times-larger ResNet-18 teacher. BA-MTKD further reduces the dense angle error by 26–29% and shows the smallest degradation under four image perturbations, and all six primary gains are significant after Holm adjustment. The deployed student has 1.00 M parameters and 0.179 G MACs and runs in 3.4 ms per image on a desktop GPU. These offline benchmark results establish BA-MTKD as a compact perception module for the next validation stage: closed-loop robotic grasping on embedded hardware.
Authors
- Bowen Xing (ORCID: https://orcid.org/0000-0002-2209-4381)
- Wugui Wang
- Bing Li (ORCID: https://orcid.org/0000-0002-2135-1892)
- Dongshi Bian
- Lei Qiao
Institutions
- Harbin Engineering University (CN)
- Wuhan Ship Development & Design Institute (CN)
- Marine Design & Research Institute of China (CN)
- Shanghai Ocean University (CN)
Publication Details
- Journal
- Electronics
- Published
- 2026-11-18
- DOI
- https://doi.org/10.3390/electronics15225379
- Primary Topic
- Robot Manipulation and Learning
- Type
- article
- Field-Weighted Citation Impact
- 0.00