Towards Model-Free Tool Dynamic Identification and Calibration Using Multi-Layer Neural Network

Area of Science:

• Robotics
• Control Systems
• Machine Learning

Background:

• Accurate interaction force sensing is crucial for precise robot control in tasks like surgery and teleoperation.
• Existing methods often struggle with tool dynamics, as force sensors measure both interaction and gravity forces, necessitating tool dynamic identification for model-based control.

Purpose of the Study:

• To propose and evaluate a model-free technique for dynamic identification of robot tool dynamics.
• To enhance the accuracy of dynamic models compared to traditional model-based methods.

Main Methods:

• Utilized multi-layer neural networks (MNN), specifically feed-forward (FF-MNN) and cascade-forward (CF-MNN) architectures, for model-free dynamic identification.
• Compared the MNN approach against a model-based curve fitting (CF) technique.

Main Results:

• The MNN-based model-free technique demonstrated improved accuracy in tool dynamic identification compared to the model-based curve fitting method.
• Demonstrated successful application in bilateral teleoperation using a KUKA LWR4+ robot and a Force Dimension SIGMA 7 manipulator.

Conclusions:

• The proposed model-free dynamic identification using MNN offers a promising advancement over model-based methods.
• This technique enhances control precision and performance in robot-assisted surgery and bilateral teleoperation systems.