Biped Gait Stability Classification Based on the Predicted Step Viability.

Area of Science:

• Robotics
• Machine Learning
• Control Systems

Background:

• Bipedal walking robots and exoskeletons require robust stability control.
• The Predicted Step Viability (PSV) algorithm offers advanced gait planning but is computationally intensive.
• Real-time implementation of complex optimization criteria like PSV is a significant challenge.

Purpose of the Study:

• To evaluate the feasibility of real-time implementation for the PSV algorithm in bipedal locomotion.
• To reduce the computational cost of PSV by using machine learning classifiers for step stability prediction.
• To assess the performance of various classification algorithms and a stacking strategy for predicting gait stability.

Main Methods:

• Generated three datasets of increasing complexity using PSV simulations.
• Trained and evaluated 11 classification algorithms (including kNN, SVM, Decision Tree, Random Forest, MLP) to predict step stability.
• Utilized a stacking strategy in conjunction with classifiers to enhance prediction accuracy.
• Measured computational cost and prediction time for each method.

Main Results:

• k Nearest Neighbors, Support Vector Machine (RBF Kernel), Decision Tree, and Random Forest showed superior performance in predicting step viability.
• Multi-Layer Perceptron also demonstrated consistent, strong performance.
• Linear-based algorithms exhibited lower predictive capabilities.
• The stacking strategy did not yield significant performance improvements.
• Classifiers achieved rapid computation (<1 ms) with minimal computational overhead.

Conclusions:

• Machine learning classifiers can effectively predict step viability for bipedal robots, significantly reducing computational cost.
• The developed feature vector and classification approach are adaptable to different robotic models and datasets.
• Sufficient and balanced training data are crucial for reliable classifier performance in gait stability prediction.
• This method enables efficient, real-time gait planning essential for stable bipedal locomotion.