Neural Network-Based Model Predictive Trajectory Tracking Control for Dual-Motor-Driven a Tracked Unmanned Vehicle

Area of Science:

• Robotics
• Control Systems Engineering
• Artificial Intelligence

Background:

• Trajectory tracking is crucial for unmanned vehicle navigation.
• Existing control models significantly impact vehicle performance.
• Dual-motor-driven tracked unmanned vehicles (TUVs) present unique control challenges due to coupled dynamics.

Purpose of the Study:

• To enhance trajectory tracking accuracy for dual-motor-driven TUVs.
• To develop a data-driven model predictive control (MPC) scheme.
• To investigate the application of neural networks in MPC for TUVs.

Main Methods:

• Developed a Long Short-Term Memory (LSTM) network for TUV dynamics modeling.
• Utilized multi-body dynamics for LSTM model training and validation.
• Designed a neural network-based MPC (NN-MPC) using the LSTM model within a receding horizon framework.
• Computed optimal motor torques for trajectory tracking.

Main Results:

• The proposed NN-MPC demonstrated superior trajectory tracking compared to a physics-model based MPC.
• Reduced root mean square (RMS) lateral error by 12.1% and heading error by 7.9% at medium speeds.
• Reduced RMS lateral error by 80% and heading error by 14.0% at high speeds.
• Field experiments confirmed the practical feasibility of the NN-MPC scheme.

Conclusions:

• The data-driven NN-MPC approach significantly improves trajectory tracking accuracy for TUVs.
• LSTM networks effectively model complex TUV dynamics for predictive control.
• The proposed method offers a viable and effective solution for TUV control, outperforming conventional strategies.