Addressing catastrophic forgetting in payload parameter identification using incremental ensemble learning.

Area of Science:

• Robotics
• Machine Learning
• Industrial Automation

Background:

• Collaborative robots (cobots) are vital in Industry 4.0 manufacturing.
• Dynamic environments necessitate rapid reconfiguration and payload identification.
• Previous methods like Incremental Ensemble Model (IEM) had limitations, including catastrophic forgetting.

Purpose of the Study:

• To develop a novel incremental ensemble learning method for cobot payload parameter identification.
• To address and eliminate the issue of catastrophic forgetting in identification models.
• To enhance the accuracy and adaptability of cobot control systems.

Main Methods:

• Introduced a new incremental ensemble learning approach by adding weak learners for each training bag.
• Developed a classification model to select the most accurate weak learner for new data.
• Implemented the method for real-time incremental updates during cobot operation.

Main Results:

• The proposed method effectively prevents catastrophic forgetting.
• Demonstrated superior accuracy and adaptability compared to prior methods.
• Successfully validated on a Franka Emika cobot, showing robust performance.

Conclusions:

• The novel incremental ensemble learning method offers a robust solution for cobot payload identification.
• This approach ensures precise and safe cobot operation in dynamic manufacturing settings.
• Eliminates catastrophic forgetting, enabling continuous learning and adaptation.