Machine Learning Models for Inverse Kinematic Prediction of a Wrist Mechanism in Robotic Charging

The presentation "Comparative Study of Machine Learning Models for Inverse Kinematic Prediction of a Flexible-Tube Wrist Mechanism in Robotic Charging Stations" addresses the challenge of predicting inverse kinematics (IK) in flexible systems caused by 6-DOF passive deflections. Various machine learning (ML) models, including MLP, RNN, LSTM, Random Forest, XGBoost, LightGBM, SVR, GP, and a hybrid physics-based approach, are evaluated using synthetic datasets that simulate end-effector motion and deflection.
Results show the hybrid model achieves perfect accuracy (RMSE = 0.00), while XGBoost performs best among ML models (RMSE = 0.11). LSTM performs well for time-dependent data (RMSE = 0.32), whereas MLP struggles with system complexity (RMSE = 2.99). The study concludes that combining physics-based insights with ML enhances prediction accuracy. Future work includes optimizing top-performing models and conducting experimental validation for real-time docking precision.