Noise Parameter Tuning for an Invariant Extended Kalman Filter Implementation Using Genetic Algorithms

Accurate state estimation is essential for robotic applications, and factors such as sensor noise and parameter tuning can significantly impact performance. This paper presents the optimization of noise parameters for an Invariant Extended Kalman Filter (InEKF) using Genetic Algorithms (GA). First, experiments were conducted with the Unitree B2 quadruped robot collecting proprioceptive data, such as IMU readings and joint measurements. Using the collected information and the ground truth trajectory provided by a motion capture system, an iterative process was implemented to simulate the performance of the InEKF, serving as the fitness function for the GA. A total of 700 iterations were performed to minimize trajectory drift and Root Mean Square Error (RMSE) for the recorded trajectory. The results show a significant RMSE reduction of 59.8% and 48.2% in the X and Y axes, respectively, while maintaining robustness over longer paths. This demonstrates the effectiveness of GA-based optimization for improving InEKF performance and highlights its potential for real-world applications requiring reliable state estimation.