How to integrate the localization to a four-wheel steering platform

[doublesorrymaker]

We encountered the following issues:

Based on the above figure, the localization displayed by NDT aligns more closely with the trajectory output by GPS. However, after passing through the EKF filter, the trajectory deviates. When the vehicle is stationary, NDT registration accurately positions the vehicle. Can this issue be improved by modifying the parameters in the EKF configuration file? My idea is to increase the pose_smoothing_steps and decrease the twist_smoothing_steps parameters. Would this approach be feasible?

[Motsu-san]

@doublesorrymaker Great for asking. Could you tell us what this issue is in more detail?
I'm not sure from these graph, is this issue that the Localization output is overall offset with GPS as ground truth or that the EKF estimated position track is different from observed position one of NDT?
And I think this discussion title doesn't match this contents. Could you modify it showing your issue directly?
ex. "How to integrate the localization to a four-wheel steering platform"

[doublesorrymaker]

As the results of GPS and RTK may have errors in absolute trajectory compared to the localization results, we only analyzed their relative trajectory accuracy. Our issue is that when using the four-wheel steering platform, the NDT results are relatively accurate, but the final result has a deviation, causing the vehicle to drift when stationary. The final position is influenced by NDT positioning, drifting until corrected to the accurate position(fig.2).

[Motsu-san]

OK. Before we start talking about the drift, I would like to ask the following.
Q1: What difference on EKF did you make?
There is a big difference between NDT pose from EKF pose. It looks around 1m. I think there must be some difference on twist or EKF parameters from default. Adding it, "using the four-wheel steering platform" seems to have influence on the accuracy of twist (vehicle velocity and angular velocity).
Q2: Is the four-wheel steering platform can be fit the state space model on Autoware? (If there is a lateral move against vehicle heading, it is not fit by default, I think.)
Q3: How much are the gain error of vehicle velocity and the bias error of angular velocity remaining after caribration?

[doublesorrymaker]

Answer 1: I did not change the relevant configuration parameters of the EKF to obtain the test results shown in the figure. I also tried to change the proc_stddev_vx_c and proc_stddev_wz_c parameters to replace them with the actual values tested by the vehicle, but the localization trajectory results did not improve.

Answer 2: I agree with this point of view. I tried to change and replace the state space model used by the EKF, but I have not tested it yet. I think this is the key to the problem.

Answer 3: I am a bit unclear about which module you are referring to for these two parameters. If you are referring to proc_stddev_vx_c and proc_stddev_wz_c under ekf_localizer, I have tested and analyzed them to be 5.8253 and 2.96191, respectively. If you are referring to velocity_stddev_xx and angular_velocity_stddev_zz under vehicle_velocity_converter, the results are 0.389375 and 0.0315789, respectively. (Regarding the angular_velocity_stddev_zz parameter, I am a bit confused about whether it is calculated using the wheel's angular velocity or the vehicle's angular velocity?)

Thank you for your patient response.