Creating Semi-supervised learning-based Adaptable Object Detection Models for Autonomous Service Robot

Page view(s)
71
Checked on Apr 16, 2024
Creating Semi-supervised learning-based Adaptable Object Detection Models for Autonomous Service Robot
Title:
Creating Semi-supervised learning-based Adaptable Object Detection Models for Autonomous Service Robot
Journal Title:
SSRN Electronic Journal
Keywords:
Publication Date:
13 April 2022
Citation:
Chang, R., Pahwa, R. S., Wang, J., Chen, L., Satini, S., Wan, K. W., & Hsu, D. (2022). Creating Semi-supervised learning-based Adaptable Object Detection Models for Autonomous Service Robot. SSRN Electronic Journal. https://doi.org/10.2139/ssrn.4075994
Abstract:
Autonomous robots are now widely used in smart factories as they have led to improvements in productivity and reduction in long-term operating costs. New smart factories have also been designed to integrate the use of robots and include static and close environment where objects and routes are completely predefined. This is a usual requirement for mobile robots since their perception capability is usually limited to a predefined number of objects. Most detection models are trained beforehand with a huge amount of annotated data and are difficult to adapt to new objects or change in existing definition or classes. Therefore, visionbased robotic navigation and path planning are also limited and any adaptation or a change of environment is often a challenge. Semi-supervised detection models have been introduced to tackle this problem. With the use of a limited amount of labelled data together with unlabelled data, models can still be trained and achieve a better performance than fully supervised models. However, they are still affected by data bias, noisy labels and are usually applied on limited datasets. In this paper, we first demonstrate how data distribution and change in robot sensors significantly reduce the accuracy of existing detection models. We discuss a preselection strategy for selecting data to be labelled and then introduce a semi-supervised approach that outperforms fully supervised approach.We successfully apply it to a real-world robot delivery scenario in a hospital environment and demonstrate that our model improves the detection accuracy compared to fully supervised methods.
License type:
Publisher Copyright
Funding Info:
This research / project is supported by the A*STAR - National Robotics Programme (NRP)
Grant Reference no. : 192 25 00049

This research / project is supported by the A*STAR - AI3 HTPO Seed Fund
Grant Reference no. : C211118008
Description:
Conference Website - https://www.clf2022.sg/programme 12th Conference on Learning Factories, CLF2022 Paper Session 4b - Machine Learning and AI in Learning Factories Time: Wednesday, 13/Apr/2022: 2:45pm - 3:00pm ID: 440 / PS-4b: 3 Final Full/Short Paper Submission Virtual attendance
ISSN:
1556-5068