How to improve the score of how well the map aligns with the sensor points?

[YamashitaYz]

Hello everyone, I am debugging the relevant parameters of ndt_scan_matcher.param.yaml. Among them, converged_param_type selects NEAREST_VOXEL_TRANSFORMATION_LIKELIHOOD, sets the threshold parameter converged_param_nearest_voxel_transformation_likelihood to 3.0, and I will upload the other parameter settings now. How can I improve the score of the degree of alignment between the map and the sensor point? Or what parameters are related to the score of the degree of alignment between the map and the sensor point?
The difficulties I have debugged are:

1. When the converged_param_nearest_voxel_transformation_likelihood threshold is set too small, after initialization, the map and the point cloud are aligned, but the point cloud soon has a misalignment phenomenon.
2. When the converged_param_nearest_voxel_transformation_likelihood threshold is set too large, after initialization, the map and the point cloud are aligned, and there is no misalignment. But the [Auto] button cannot be clicked
3. When the converged_param_nearest_voxel_transformation_likelihood threshold is set to balance between the two, after initialization, the map and point cloud are aligned, and there is no misalignment. The [Auto] button can be clicked normally. After [rviz2] publishes the point [2D Goal Pose], the point cloud will be misaligned, causing the vehicle to be unable to move.
4. I found that the value of resolution will affect the score when there is no [INITIALIZED], but I don’t know how to debug it.
   Can you share your debugging experience? Thank you!
   My ndt_scan_matcher.param.yaml content is:

/**:
  ros__parameters:
    frame:
      # Vehicle reference frame
      base_frame: "base_link"

      # NDT reference frame
      ndt_base_frame: "ndt_base_link"

      # Map frame
      map_frame: "map"


    sensor_points:
      # Required distance of input sensor points. [m]
      # If the max distance of input sensor points is lower than this value, the scan matching will not be performed.  #tues
      required_distance: 10.0


    ndt:
      # The maximum difference between two consecutive
      # transformations in order to consider convergence
      trans_epsilon: 0.01

      # The newton line search maximum step length
      step_size: 0.1 #tues

      # The ND voxel grid resolution #tues
      resolution: 2.0

      # The number of iterations required to calculate alignment
      max_iterations: 30

      # Number of threads used for parallel computing #tues
      num_threads: 4

      regularization:
        enable: false

        # Regularization scale factor #tues
        scale_factor: 0.01


    initial_pose_estimation:
      # The number of particles to estimate initial pose #tues
      particles_num: 200

      # The number of initial random trials in the TPE (Tree-Structured Parzen Estimator).
      # This value should be equal to or less than 'initial_estimate_particles_num' and more than 0.
      # If it is equal to 'initial_estimate_particles_num', the search will be the same as a full random search.
      n_startup_trials: 20


    validation:
      # Tolerance of timestamp difference between current time and sensor pointcloud. [sec]
      lidar_topic_timeout_sec: 1.0

      # Tolerance of timestamp difference between initial_pose and sensor pointcloud. [sec]
      initial_pose_timeout_sec: 1.0

      # Tolerance of distance difference between two initial poses used for linear interpolation. [m] #tues
      initial_pose_distance_tolerance_m: 10.0

      # The execution time which means probably NDT cannot matches scans properly. [ms]
      critical_upper_bound_exe_time_ms: 100.0


    score_estimation:
      # Converged param type
      # 0=TRANSFORM_PROBABILITY, 1=NEAREST_VOXEL_TRANSFORMATION_LIKELIHOOD
      converged_param_type: 1

      # If converged_param_type is 0
      # Threshold for deciding whether to trust the estimation result #tues
      converged_param_transform_probability: 1.8

      # If converged_param_type is 1
      # Threshold for deciding whether to trust the estimation result #tues
      converged_param_nearest_voxel_transformation_likelihood:  2.5

      # Scan matching score based on no ground LiDAR scan
      no_ground_points:
        enable: false

        # If lidar_point.z - base_link.z <= this threshold , the point will be removed #tues
        z_margin_for_ground_removal: -0.2


    covariance:
      # The covariance of output pose
      # Note that this covariance matrix is empirically derived
      output_pose_covariance:
        [
          0.0225, 0.0,   0.0,   0.0,      0.0,      0.0,
          0.0,   0.0225, 0.0,   0.0,      0.0,      0.0,
          0.0,   0.0,   0.0225, 0.0,      0.0,      0.0,
          0.0,   0.0,   0.0,   0.000625, 0.0,      0.0,
          0.0,   0.0,   0.0,   0.0,      0.000625, 0.0,
          0.0,   0.0,   0.0,   0.0,      0.0,      0.000625,
        ]

      # 2D Real-time covariance estimation with multiple searches (output_pose_covariance is the minimum value)
      covariance_estimation:
        enable: false

        # Offset arrangement in covariance estimation [m]
        # initial_pose_offset_model_x & initial_pose_offset_model_y must have the same number of elements.
        initial_pose_offset_model_x: [0.0, 0.0, 0.5, -0.5, 1.0, -1.0]
        initial_pose_offset_model_y: [0.5, -0.5, 0.0, 0.0, 0.0, 0.0]


    dynamic_map_loading:
      # Dynamic map loading distance
      update_distance: 20.0

      # Dynamic map loading loading radius
      map_radius: 150.0

      # Radius of input LiDAR range (used for diagnostics of dynamic map loading)
      lidar_radius: 100.0