GoKawiilFusionCore
ROS 2 sensor fusion: IMU + wheel encoders + GPS fused via UKF at 100 Hz. 22-state filter with IMU bias estimation, adaptive noise covariance, and chi-squared outlier rejection on every sensor.
Why I built this
I needed sensor fusion for a mobile robot project and reached for robot_localization like everyone does. It works well. But I wanted a filter that estimated IMU gyro and accelerometer bias as part of the state vector, adapted its noise covariance from real sensor behavior rather than config values, and rejected outliers on every sensor update: not just GPS.
So I built FusionCore. It's a 22-state UKF that fuses IMU, wheel encoders, and GPS natively. Gyro and accelerometer bias are estimated continuously as filter states. Noise covariance adapts from the innovation sequence automatically. Every sensor update: IMU, wheel odometry, GPS: goes through a chi-squared gate before it touches the filter. GPS is handled in ECEF directly, no coordinate projection.
Benchmark
FusionCore vs robot_localization on the NCLT dataset: same IMU + wheel odometry + GPS, no manual tuning. Six sequences:
RL-EKF run with odom0_twist_rejection_threshold: 4.03 and odom1_pose_rejection_threshold: 3.72 (chi²-equivalent to FusionCore's thresholds at 99.9% confidence).
Sequence FC ATE RMSE RL-EKF ATE RMSE RL-UKF 2012-01-08 5.6 m 13.0 m NaN divergence at t=31 s 2012-02-04 9.7 m 19.1 m NaN divergence at t=22 s 2012-03-31 4.2 m 54.3 m NaN divergence at t=18 s 2012-08-20 7.5 m 24.1 m NaN divergence 2012-11-04 28.6 m 9.6 m NaN divergence 2013-02-23 4.1 m 11.0 m NaN divergence
Install
... continue reading