Hardness Recognition of Robotic Forearm Based on Semi-supervised Generative Adversarial Networks

Area of Science:

• Robotics
• Machine Learning
• Tactile Sensing

Background:

• Hardness recognition is crucial for tactile sensing and robotic control.
• Deep learning methods for hardness recognition require extensive manually labeled data, increasing time and labor costs.

Purpose of the Study:

• To propose a semi-supervised generative adversarial network (GAN) to reduce the reliance on manually labeled samples for hardness recognition.
• To improve the efficiency and reduce the cost of training deep neural networks for tactile sensing.

Main Methods:

• Utilized unsupervised training of a GAN on a large unlabeled dataset to establish a good initial model state.
• Employed manually labeled samples to train the GAN, augmenting data with the trained generator.
• Pretrained and fine-tuned a hardness recognition network (HRN) using augmented and manually labeled data, inheriting architecture from the GAN's discriminator.

Main Results:

• The proposed semi-supervised GAN method significantly reduces the requirement for manual data labeling.
• Achieved excellent recognition precision in hardness recognition tasks.
• Enabled online hardness recognition by importing tactile data from robotic forearms into the trained HRN.

Conclusions:

• The developed semi-supervised GAN approach offers a viable solution for efficient and precise hardness recognition in robotics.
• This method substantially alleviates the burden of manual data annotation in deep learning models for tactile sensing.