Pose Error Prediction, Compensation Method, and Applicable Condition Determination of Parallel Motion Platform Based...

Area of Science:

• Robotics
• Machine Learning
• Control Systems

Background:

• Collecting robot configuration data is costly and time-consuming, limiting neural network applications.
• Traditional neural networks struggle with low accuracy and generalization on small datasets for robot error prediction.

Purpose of the Study:

• To develop a transfer learning approach for robot error prediction and compensation using prior kinematic knowledge.
• To propose a method for evaluating transfer learning applicability based on task similarity and data sample size.

Main Methods:

• Established a 'transfer network' integrating ideal kinematic model characteristics with actual pose data.
• Compared transfer network performance against traditional back propagation (BP) neural networks.
• Developed a method to assess transfer learning applicability by analyzing task similarity and actual pose sample numbers.

Main Results:

• The transfer network demonstrated superior performance over traditional BP networks.
• Effectively addressed low prediction accuracy and weak generalization with small-sample data.
• The proposed method accurately determined transfer learning applicability.

Conclusions:

• Transfer learning offers significant advantages for robot precision compensation, especially with limited data.
• The developed method effectively predicts the success of transfer learning in robotic applications.