[Paper Review] Lightweight Convolutional Neural Network with Gaussian-based Grasping Representation for Robotic Grasping Detection

Hu Cao, Guang Chen+3 more|arXiv (Cornell University)|Jan 25, 2021

Robot Manipulation and Learning38 references20 citations

TL;DR

This paper proposes a lightweight, fully convolutional neural network with Gaussian-based grasping representation for efficient robotic grasping detection. By encoding grasp confidence via 2D Gaussian kernels, integrating a receptive field block for multi-scale feature extraction, and employing multi-dimensional attention fusion, the model achieves 98.9% accuracy on Cornell and 95.6% on Jacquard datasets with only 4.67 million parameters—significantly smaller and faster than prior methods while maintaining state-of-the-art performance.

ABSTRACT

The method of deep learning has achieved excellent results in improving the performance of robotic grasping detection. However, the deep learning methods used in general object detection are not suitable for robotic grasping detection. Current modern object detectors are difficult to strike a balance between high accuracy and fast inference speed. In this paper, we present an efficient and robust fully convolutional neural network model to perform robotic grasping pose estimation from an n-channel input image of the real grasping scene. The proposed network is a lightweight generative architecture for grasping detection in one stage. Specifically, a grasping representation based on Gaussian kernel is introduced to encode training samples, which embodies the principle of maximum central point grasping confidence. Meanwhile, to extract multi-scale information and enhance the feature discriminability, a receptive field block (RFB) is assembled to the bottleneck of our grasping detection architecture. Besides, pixel attention and channel attention are combined to automatically learn to focus on fusing context information of varying shapes and sizes by suppressing the noise feature and highlighting the grasping object feature. Extensive experiments on two public grasping datasets, Cornell and Jacquard demonstrate the state-of-the-art performance of our method in balancing accuracy and inference speed. The network is an order of magnitude smaller than other excellent algorithms while achieving better performance with an accuracy of 98.9$\%$ and 95.6$\%$ on the Cornell and Jacquard datasets, respectively.

Motivation & Objective

• To address the trade-off between inference speed and accuracy in robotic grasping detection.
• To develop a lightweight, one-stage network suitable for real-time deployment on resource-constrained robotic systems.
• To improve feature discriminability and robustness in cluttered scenes through enhanced feature extraction and attention mechanisms.
• To encode grasp quality more effectively by emphasizing the central point with maximum grasp confidence using Gaussian kernel representation.
• To reduce network size while maintaining or improving detection accuracy compared to existing methods.

Proposed method

• A Gaussian-based grasping representation (GGR) is introduced to encode training samples, emphasizing the central point with the highest grasp confidence using a 2D Gaussian kernel.
• A receptive field block (RFB) is integrated into the bottleneck layer to enhance multi-scale feature extraction and improve feature discriminability.
• A multi-dimensional attention fusion network (MDAFN) combines pixel and channel attention to suppress noise and highlight relevant features across varying shapes and sizes.
• Low-level and high-level features are fused in the decoder path to reduce information loss during feature reconstruction.
• The network is a fully convolutional, one-stage architecture designed for end-to-end grasping pose estimation from RGB-D images.
• The model is trained on Cornell and Jacquard datasets using image-wise and object-wise splits to evaluate generalization and robustness.

Experimental results

Research questions

• RQ1 Can a lightweight, one-stage fully convolutional network achieve high accuracy and fast inference speed in robotic grasping detection?
• RQ2 How does Gaussian-based grasp representation improve the modeling of grasp confidence compared to anchor-based or keypoint-based methods?
• RQ3 To what extent does the integration of a receptive field block and multi-dimensional attention mechanism enhance feature representation and detection robustness?
• RQ4 How does the proposed method generalize to cluttered scenes with multiple objects, despite being trained on single-object datasets?
• RQ5 What is the optimal scale factor for the Gaussian kernel in terms of balancing label overlap and model stability?

Key findings

• The proposed method achieves 98.9% grasp detection accuracy on the Cornell dataset and 95.6% on the Jacquard dataset, outperforming state-of-the-art methods in accuracy-speed trade-off.
• The network size is only 4.67 million parameters, an order of magnitude smaller than competing methods such as Chu et al. (216M) and Pinto & Gupta (60M).
• The model achieves a real-time inference speed of 5ms per image, significantly faster than methods like Lenz (13.5s) and Morrison (3ms) with much smaller parameter count.
• Ablation studies confirm that all components—GGR, RFBM, and MDAFN—contribute to performance, with the full model achieving the highest accuracy (97.8% on Cornell in image-wise split).
• The model generalizes well to cluttered scenes with multiple objects, successfully detecting grasp poses even when trained on single-object data.
• Failure cases are primarily observed for small, complex-shaped, or occluded objects, suggesting that dataset diversity could further improve robustness.