Core Concepts
Neural network architectures can accurately predict joint torques of a 6-DOF robotic arm by learning the complex nonlinear dynamics, without relying on an analytical model.
Abstract
The paper presents an approach to predict the joint torques of a 6-DOF robotic arm using neural network (NN) models, without relying on traditional model-based methods that require precise knowledge of system parameters.
The key highlights are:
Three NN architectures were investigated:
Single fully connected NN
Multiple NNs for independent joint groups
Cascade NN considering joint dependencies
Data preprocessing through standardization was found to improve the performance of all NN models.
Hyperparameter optimization using Bayesian optimization further improved the accuracy, with the cascade NN model performing the best.
The cascade NN architecture, which encodes prior knowledge about joint dependencies, achieved the lowest test mean squared error (MSE) of 2.111828e-04 after hyperparameter tuning.
The results demonstrate that well-designed NN models can effectively capture the complex nonlinear dynamics, including effects like friction and joint flexibility, without requiring an analytical model of the system.
The authors suggest further research on physics-informed learning methods and strategies to reduce dataset size while maintaining accuracy, to improve the efficiency and generalization of the dynamic models.
Stats
The mean squared error (MSE) of the joint torque predictions are:
Single NN: 2.376151e-04
Multiple NNs: 2.418988e-04
Cascade NN: 2.111828e-04
Quotes
"The cascade NN architecture, which encodes prior knowledge about joint dependencies, achieved the lowest test mean squared error (MSE) of 2.111828e-04 after hyperparameter tuning."