Research On Feature Alignment Based Grasping Point Recognition Method In Robot-assisted Undressing

The global aging population,labor shortage and other social issues are driving the rapid development of assistive robots.There are various types of service-oriented intelligent robots for the daily life of users,such as rehabilitation therapy robots,communication and companionship robots,as well as assistive dressing robots.Robot-assisted dressing and undressing has become a hot research topic due to its high practicality in daily life.However,research on assistive undressing has not yet been fully carried out.This study proposes to use humanoid robots to assist healthcare workers in removing medical clothing,and this study can help alleviate the work and stress of healthcare workers to some extent.In addition,this study can also be further extended to the home environment in the future to help the elderly and disabled people to provide various types of clothing removal services,which has high practical value and application prospects.By giving the robot a variety of perceptual capabilities,the researchers enable it to manipulate non-rigid,high-dimensional clothing in close human-robot interaction and thus realize the process of assisting users in dressing.The core problem of this study is to train a vision model to provide the robot with grasping position information about the user’s shoulders.In fact,the training of vision models usually relies on high-quality labeled data,and therefore,a large amount of real image data needs to be collected and labeled in the preliminary work.However,the acquisition of real data with labels is both time-consuming and difficult to obtain.Compared to real data,synthetic data,which can be acquired easily in large quantities,can effectively augment the hard-to-capture real data.However,the natural domain gap between real data and synthetic data(e.g.data distribution,etc.)significantly weaken the performance improvement of synthetic data for models.To overcome the limitation,this paper proposes a Syn-to-Real Alignment Network,SRAN.In terms of network design,the alignment network not only adaptively aligns synthetic and real images at the feature level using Deformable Convolution to bridge the domain gap,but also generates reasonably distributed pseudo-labels for the aligned images through the Mixup method with the convolution module.The aligned images generated by the alignment network are then applied to the convolutional neural network-based grasping point recognition model for training.In the evaluation of model performance,it is verified that the alignment network has a significant role in reducing domain gap and improving the accuracy of recognizing grasping points.Furthermore,in order to deploy the grasping point model for validation in real experiments,a robot-assisted undressing system is designed in this thesis.The system first classifies the clothes,then uses the grasping point model to identify the grasping points of the real-time image and finally the robot completes the overall undressing work through path planning.Real-time experiments show that the proposed undressing system can enable a robot to classify the type of clothes and recognize the grasping points to provide the undressing assistance for real users.