Problems with 3D local SLAM

[halops]

Hi, I am having problems with the 3D local SLAM tuning. I am using 2x VLP-16 and a Microstrain 3dm-gx5-45. The main issue that I have is that the local trajectory slightly oscillates, as you can see in the picture below. My configuration is the following:

options = { map_builder = MAP_BUILDER, trajectory_builder = TRAJECTORY_BUILDER, map_frame = "map", tracking_frame = "microstrain_imu_link", published_frame = "base_link", odom_frame = "odom", provide_odom_frame = true, publish_frame_projected_to_2d = false, use_odometry = false, use_nav_sat = false, use_landmarks = false, num_laser_scans = 0, num_multi_echo_laser_scans = 0, num_subdivisions_per_laser_scan = 1, num_point_clouds = 2, lookup_transform_timeout_sec = 0.2, submap_publish_period_sec = 0.3, pose_publish_period_sec = 5e-3, trajectory_publish_period_sec = 30e-3, rangefinder_sampling_ratio = 1., odometry_sampling_ratio = 1., fixed_frame_pose_sampling_ratio = 1., imu_sampling_ratio = 1., landmarks_sampling_ratio = 1., }

--TRAJECTORY BUILDER PARAMS --Providing full 360 scans with two lidars (and per-point stamp) TRAJECTORY_BUILDER_3D.num_accumulated_range_data = 6

--Setting up a voxel filter to reduce the density of the data VOXEL_SIZE = 0.05 TRAJECTORY_BUILDER_3D.voxel_filter_size = VOXEL_SIZE

MAP_BUILDER.use_trajectory_builder_3d = true MAP_BUILDER.num_background_threads = 12 POSE_GRAPH.optimization_problem.huber_scale = 5e2 POSE_GRAPH.optimize_every_n_nodes = 200 POSE_GRAPH.constraint_builder.sampling_ratio = 0.03 POSE_GRAPH.optimization_problem.ceres_solver_options.max_num_iterations = 20 POSE_GRAPH.constraint_builder.min_score = 0.62 POSE_GRAPH.constraint_builder.global_localization_min_score = 0.66

Changing the num_accumulated_range_data from 2 to 6 helped reduce the oscillation, but it is still present. Thanks for your help!

[tyhowell]

Sounds like a tuning question - if you are using the ROS velodyne driver have you tried setting npackets to 1? That would provide ~75 packets per rotation to Cartographer, so you would need to increase _TRAJECTORY_BUILDER_3D.num_accumulated_rangedata appropriately, the original authors had it set to 160 for two velodynes.

[halops]

I am working offline with recorded data and I can't test this right now, I will try this as soon as I can. However, I fail to understand how should this produce a different effect to the mapping, since it should be equivalent to providing full 360º scans with per-point timestamp, right? Thanks for your help!