Getting outputs for multiple robots

[TakShimoda]

Hello, thank you for helping me solve #43 which was for a single robot case. I can get optimized poses as g2o files and the pose timestamps, and can combine them to create a TUM file to evaluate ATE and RPE with evo.

I am right now stuck on getting proper pose outputs for two robots. My terminals are: ros2 launch cslam_experiments rtabmap_kitti_stereo_odometry.launch.py frame_id:=B01_link namespace:=/r0 args:=--delete_db_on_start use_sim_time:=false odom_topic:=B01_odom ros2 launch cslam_experiments rtabmap_kitti_stereo_odometry.launch.py frame_id:=B04_link namespace:=/r1 args:=--delete_db_on_start use_sim_time:=false odom_topic:=B04_odom

• Frontends with specified frame_id and odom_topic for robots named B01 and B04.

ros2 launch cslam_experiments cslam_stereo.launch.py namespace:=/r0 robot_id:=0 max_nb_robots:=2 config_file:=B01_stereo.yaml ros2 launch cslam_experiments cslam_stereo.launch.py namespace:=/r1 robot_id:=1 max_nb_robots:=2 config_file:=B04_stereo.yaml

• Launching robots with the namespacing and specific config files, which are for the TF frame names and topics.

Then I play the bag files with renamed topics and qos override just for /tf_static.

Here is what works and doesn't work:

• If I just launch rtabmap_kitti_stereo_odometry.launch.py and cslam_stereo.launch.py for just one of the robots (r0 and r1) and play back the appropriate bag file, it works (i.e. can see visualization and proper outputs to log folder)
• If I launch rtabmap_kitti_stereo_odometry.launch.py and cslam_stereo.launch.py for both robots and play back the bag file for r0, it works.
• However, if I launch rtabmap_kitti_stereo_odometry.launch.py and cslam_stereo.launch.py for both robots and play back the bag file for r1, then it doesn't work, as there are no outputs. Note the main log folder is there, but it doesn't create the subfolders every 10 seconds as in the working case.

I tried combing through the source code, and here's where I can trace the issue:

• The callback to write to logs is here, which subscribes to /<namespace>/cslam/optimized_estimates
• The function share_optimized_estimates publishes to the topic above.
• shared_optimized_estimates is called from check_result_and_finish_optimization, which is run here
• So I notice the variable optimizer_state_ wasn't what it should be to trigger the functions mentioned above. When checking on the correctly working examples, using some logging, their states would be: waiting->POSEGRAPH_COLLECTION->START_OPTIMIZATION-> OPTIMIZATION
• current_neighbors_callback changes the optimizer_state_ , by checking the is_optimizer() function. Here is where I check why the working examples work but the specific example I mentioned above fails

Here are the different cases of what the parameters are in current_neighbors_callback and the is_optimizer() functions

• In the case of just launch files for one robot (either r0 or r1), in current_neighbors_callback, the *msg->origins.ids.size() = 0, hence this loop never triggers, so is_optimizer is never set to false (I also checked odometry_pose_estimates_ and it's always increasing in size for every use case).
• For r0 and r1 nodes running, and bag files played for just one of them, here is a table summarizing the parameters and the result:

Parameters for robot #	msg->robots	msg->origins	msg->origins.ids.size()	msg->origins.ids[0]	msg->origins.ids[0]	origin_robot_id_	robot_id_	is_optimizer
0	1747769088	1747769120	1	1	1	0	0	true
1	-1639485936	-1639485904	1	0	0	1	1	false, because origin_robot_id_ > msg->origins.ids[0]

So from this analysis, it simply looks like it's designed so only the lowest robot_id can output poses. How would I work around this? I tried all kinds of different combinations, such as having all robot_id the same but it gives an error in the neighbors_manager.py script. Obviously, other things, such as making robot_id negative also gives errors.

I was also trying to debug the topic cslam/get_current_neighbors that current_neighbors_callback subscribes to, and is published to here but it seems confusing as it seems to publish an initialized std_msgs::msg::String() so I can't seem to debug its value and understand what exactly is published on the topic.

I was wondering how I could get the system to work properly so I can get pose outputs for all robots.

Thanks,

[lajoiepy]

Hi! Indeed, only the robot with the lowest ID produces results, but the results will include the estimates from all robots (i.e. the merged map of all robots). The robots need to have different positive IDs. cslam/get_current_neighbors is simply a request for the the neighbors list, hence the empty message. Once request received, it will be published here on the topic cslam/current_neighbors Importantly to merge the maps, Swarm-SLAM needs to find loop closures between the maps. If there are some inter-robot loop closures found, you should see the message New inter-robot loop closure measurement printed from here.

[TakShimoda]

Hi, I do notice I get outputs for robot 0, and it seems to involve all robots. I get loop closure measurement outputs in the terminal, and the visualization does show the pose graphs being updated.

The outputs in the log folder for 3 robots looks like this:

• The gps files are empty because I'm not using gps
• spectral_matches.csv is empty, although I feel maybe it shouldn't be?
• The optimized_global_pose_graph.g2o file has more rows for VERTEX_SE3:QUAT than rows in pose_timestamps0.csv, where they usually have the same number of rows in the single robot case

If I match the all the timestamps in pose_timestamps0.csv (e.g. x) to the first x rows of VERTEX_SE3:QUAT I get the resulting trajectory: Although the actual trajectories are 3 concentric circles of differing radius:

So from here, my main issues are:

• the results are poor (it could be my specific bag files, but I doubt it as single-robot odometry from RTABMAP stereo is highly accurate, even without inputting intrinsic/extrinsic parameters)
• As shown from my output trajectory, I can't discern each individual robots' trajectory, although I want to do that (assess ATE/RPE for each of the 3 robots). I know from here that the large numbers in the second column of the g2o files are gtsam labels, but I can't see how I can derive robot ID's from there, and I don't know how to relate the timesstamps of the pose to pose_timestamps0.csv (in this csv file, I noticed the first column vertice_id is all the same which isn't a problem in single robots, but is an issue for multi-robots)

So I wanted to know how I can get the individual robot trajectories and timestamps, and also see how to improve the estimates as well? From figure 4 of your paper, I can see individual robot trajectories:

[TakShimoda]

I just checked, and I realized I actually did get outputs for robots 1 and 2. What I noticed is for robot 1, I get the usual output, although it was fairly inaccurate. Meanwhile for robot 2, all I get for the outputs are log.csv and pose_timestamps2.csv.

So now it seems: robot 0 folder outputs all combined poses, robot 1 folder outputs its own poses, and robot 2 doesn't output its poses (possibly because it wasn't involved in an inter-robot loop closure).

[lajoiepy]

Here are some functions to reverse the GTSAM id format and convert g2o files to TUM for evo https://gist.github.com/lajoiepy/10fb6a09ec23d09a85166bf8a641c7ee

As for the accuracy, it depends on many parameters. Swarm-SLAM is a framework that puts together many third party libraries, and each has its own parameters that need tuning.

For example, in your case, you should look into the Visual Place Recognition (default: https://github.com/gmberton/CosPlace) and RTAB-Map for the 3D registration. I exposed many params for those techniques in the config files, but if some that you need are not exposed, I suggest that you change them directly into the source code. For example, for RTAB-Map I expose the minimum number of keypoints inliers for the registration. Other parameters can be changed here https://github.com/lajoiepy/cslam/blob/32527430ef271064d468f442a3ab754917abb011/src/front_end/rgbd_handler.cpp#L118

Unfortunately, figuring out the right parameters for each components in order to produce high quality loop closures on your dataset can be a tedious process. Let me know if something is unclear.

[TakShimoda]

Hi, sorry for the delayed response.

Thank you for the python script, it helped me create separate tum files for each robot.

As for the accuracy, outside of tweaking the parameters, I noticed that CosPlace only uses RGB images as mentioned here, but my dataset is only in monochrome images. So I recorded both stereo and RGB images on the same camera (d435i), and fed stereo to the odometry and RGB-D to the loop closure node. I copied the rgbd yaml here to pass to the loop detection, map manger, and pose graph manger nodes, except I added a sensor_base_frame_id field similar to the stereo yaml since it didn't have that and my frame link isn't "D435i_link"

In this case, I'm assuming I'm doing this correctly? I play back both stereo and RGB data and the loop closure node will get the RGB topic from the config yaml. I realized my depth image scale (320x240) is different from RGB (640-480) so it gave these messages:

[map_manager-2] [WARN] [1724081254.103768736] [cslam]: Depth image size (320x240) doesn't match RGB image size (640x480)!
[map_manager-2] Failed to find match for field 'x'.
[map_manager-2] Failed to find match for field 'y'.
[map_manager-2] Failed to find match for field 'z'.
[map_manager-2] Failed to find match for field 'rgb'.

Nevertheless, it did manage to close loops for intra-robot loop closure, and still recognized inter-robot loop closure but failed to transmit them so far (it gave a lot of messages mentioned failed inter-robot loop closure measurements). So I probably need to take another experiment with depth images of the same same dimension and more overlap in robot trajectories.

I had a few questions though:

• Since there's a constant offset between RGB and stereo frames, do I have to account for this by providing extrinsic values somewhere?
• Does the loop closure need depth image? I didn't see any rgb yaml files to provide as config and if I tried to set sensor_type in the yaml file to rgb, it says it's not supported: [r1.cslam_map_manager]: Sensor type not supported: rgb

[lajoiepy]

1- In the current implementation, it is assumed that the loop closing and odometry sensors are the same. For optimal accuracy, you should add this offset (extrinsics) in the loop closure transforms. However, pose graph optimization is fairly robust to small errors (such as the distance between the stereo and rgb sensors on the realsense), so you will not see much of a difference in terms of accuracy. 2- Yes the loop closures need depth or stereo, otherwise we can't get the scale information. So in your case, you could feed the stereo pairs corresponding (in time) to the RGB image used for loop closure detection.

[TakShimoda]

Okay, I'll test with RGB-D loop closure again this time with the same color and RGB image sizes, and I'll try tweaking some of the threshold parameters. Tracking is fairly accurate and intra-robot loop closure works well, so I'll try to see if I can adjust some CosPlace parameters to improve the merging process for different robots. Thanks!