GGCNN Model Improvements for Low Computing Resource Environments

Advisor

Team

Abstract

Despite the effectiveness of deep neural networks in robotic grasp detection, deploying them on resource-constrained platforms remains challenging due to high computational demands. This study proposed a knowledge distillation-based approach to compress the Generative Grasping Convolutional Neural Network (GGCNN), enabling efficient real-time performance without compromising grasping accuracy. Two lightweight student models, Version 1 and Version 2, are designed using distinct distillation strategies to strike a balance between model size, inference speed, and accuracy. Experimental results show that both student models significantly reduce model size and inference time, with one achieving up to a 75% size reduction and nearly halving the inference time, while maintaining competitive IoU accuracy. Notably, Version 2 matches the teacher model’s accuracy while offering improved efficiency and higher throughput, demonstrating the effectiveness of the method for real-time robotic applications where speed and resource efficiency are essential.