Hybrid vehicle state estimation using closed-form liquid neural networks and nonlinear Kalman filtering

Area of Science:

• Robotics and Control Systems
• Artificial Intelligence and Machine Learning
• Automotive Engineering

Background:

• Accurate vehicle dynamic state estimation is vital for advanced driver-assistance systems and autonomous driving.
• Direct measurement of critical states like sideslip angle is challenging due to sensor limitations and noise.

Purpose of the Study:

• To propose a novel hybrid framework for robust vehicle state estimation.
• To enhance the accuracy and reliability of estimating vehicle states, particularly sideslip angle, under adverse conditions.

Main Methods:

• Integration of closed-form continuous-time (CfC) networks with an enhanced unscented Kalman filter (UKF).
• Utilizing CfC networks for sequential IMU and GPS data processing to predict states and uncertainties.
• Incorporating CfC predictions as pseudo-measurements in the UKF with uncertainty-based measurement covariance.
• Developing a data-driven sigma points sampling strategy for improved observation distribution approximation.

Main Results:

• The hybrid framework demonstrated superior accuracy and robustness compared to baseline methods on both simulated (KITTI, Carsim) and real-world data.
• The proposed method effectively handles unknown noise, improving vehicle state estimation reliability.
• Validation through extensive experiments confirms the framework's effectiveness.

Conclusions:

• The hybrid CfC-UKF framework offers a significant advancement in vehicle state estimation.
• This approach enhances the safety and performance of autonomous driving systems by providing reliable state predictions.
• The open-source code facilitates further research and development in the field.