MultivariateSystem Identification of Differential Drive Robot: Comparison Between State-Space and LSTM-Based Models

Area of Science:

• Robotics
• Machine Learning
• Control Systems

Background:

• Accurate mobile robot modeling is essential for odometry estimation, control, and navigation.
• Classical state-space models (SSMs) and advanced deep learning models like Long Short-Term Memory (LSTM) networks are used for system identification.
• Direct comparisons between SSMs and LSTMs for mobile robot odometry are limited.

Purpose of the Study:

• To compare the performance of a classical SSM and an LSTM-based recurrent neural network for modeling a differential drive robot.
• To evaluate prediction accuracy, odometry estimation, and computational efficiency of both modeling approaches.
• To provide insights into the suitability of each method for mobile robot applications.

Main Methods:

• A differential drive robot was modeled using both a classical SSM and an LSTM recurrent neural network.
• Experimental data from a five-minute navigation sequence was used to train and test the models.
• Performance was assessed based on prediction accuracy (FIT), odometry estimation error (RMSE), and computational resource usage.
• Both models underwent tuning: order selection for SSM and hyperparameter optimization for LSTM.

Main Results:

• The optimal SSM was a second-order model, while the LSTM utilized seven layers and 30 neurons.
• The SSM achieved higher prediction accuracy (FIT: 94.70% for linear, 91.71% for angular velocity) and lower odometry RMSE (0.85 m, 0.17 rad) compared to the LSTM (FIT: 93.10%, 90.95%; RMSE: 1.09 m, 0.23 rad).
• The SSM demonstrated significantly greater computational efficiency (0.00257 ms/step, 1.03 bytes/step) than the LSTM (0.0342 ms/step, 20.49 bytes/step).

Conclusions:

• Classical SSMs remain a highly effective and computationally efficient choice for accurate mobile robot odometry estimation.
• LSTM networks show flexibility through tuning but were outperformed by SSMs in this specific comparison.
• Future research could explore hybrid models combining the strengths of both approaches for enhanced robustness in complex environments.