Qustion about Integrating SLAM (Point-LIO)

[rain1204]

Hi,I’m considering using a SLAM module like Point-LIO with this planner. Would integrating it be as simple as replacing the cloud_in topic and state_estimation with cloud_registered and Odometry provided by the SLAM module? Or would additional modifications be required?

Thanks in advance!

[caoyuhong001]

That technically should work. In my experiment, I used the CMU development environment (https://www.cmu-exploration.com/development-environment) for waypoint following, which takes the output of a SLAM module (I used FAST-LIO2) as the input. Therefore, in the default setup, the ARiADNE planner gets the cloud and odometry from that module, but this is flexible.

[rain1204]

Hello, thank you very much for your previous help. I was able to use the PointLIO results for exploration. However, I'm currently facing an issue: when the robot is moving, there are dynamic objects nearby (e.g., people walking), and the point cloud from those moving people gets inserted into the octomap. This causes the exploration to end prematurely, as those dynamic obstacles are treated as static.

Is there any recommended way to handle this situation? Thank you!

[caoyuhong001]

The octomap has some parameters for the occupied probability. I believe these two are the most important for dynamic objects.

    <param name="sensor_model/hit" value="0.8"/>
    <param name="sensor_model/miss" value="0.45"/>

I used these values in my experiment, working well with cars and people.

By the way, the code I open-sourced in this repo may not work perfectly in a real-world, large-scale environment, as I directly classify the free/obstacle through the z value. In practice, it would be better to have a traversability evaluation module (this is what I used in my following work, but I am afraid I can not share it before our new paper gets accepted).

[rain1204]

Thank you for your reply ! I will test these parameters. Just to confirm, when you mention classifying free space and obstacles through the z value, do you mean that the distinction is based purely on the height of the points? Also, I’m curious—what kind of real-world environments do you think this repo is best suited for?

[caoyuhong001]

Yes, it is only based on the height, you can check Octomap's code. If you want to directly use this repo, you will need to align the z axis with the gravity first, then make sure your environment is flat. If not you may find the ground classified as obstacle.

I would like to suggest you try to implement the traversability analysis module, so this planner can work in uneven terrain. With that function, I have recently achieved successful exploration in our campus like this:

[rain1204]

Wow, looks great!
I have a couple of questions:

1. Does the traversability analysis module essentially replace the role of Octomap in your system?

2. During campus exploration, were there any areas that you wanted the robot to avoid? If so, is there a way to set constraints or restricted zones in the planner?

By the way, I've tested those parameters and they work !

Thank you so much for your suggestions!

[caoyuhong001]

Glad to hear those parameters work for you!

The traversability analysis module I implemented is based on the CMU exploration repo:https://github.com/HongbiaoZ/autonomous_exploration_development_environment/blob/noetic/src/terrain_analysis_ext/src/terrainAnalysisExt.cpp, which classifies the point cloud. I input those point clouds into two different Octomaps, one for free voxels and one for obstacle voxels. Then I project them into 2D maps and merge them into one map.

For the environment, it will be best if there are no "cliffs" in the environment, such as stairs. I found them not trivial to handle. Luckily, there are no such stairs in our campus environment, and I set a larger map resolution (1m) to avoid the robot going into the buildings. When I ran the experiment in the building, I manually put obstacles to block the stairs. Other than this, I also set a bounding box, outside of which I directly put obstacles in the map.

Btw, I asked GPT to generate functions above, and it actually worked fine. You can try it as well :-), although you may need to help GPT a little bit.

[rain1204]

Okay, I’ll look into how to integrate the traversability analysis module. Also, I’m curious about the bounding box you mentioned—how did you add it? Was it added into the Octomap during the exploration process?

In my case, the environment is an indoor circular space with a ring of floor-to-ceiling glass windows in the middle. The LiDAR point cloud tends to pass through the glass, making that area is free , which I’d like to avoid. I'm currently trying to solve this from the point cloud filtering side, but I think adding a bounding box over the glass area might also be a good solution.

[caoyuhong001]

I added the bounding box to the 2D map. If you know the relative position between the map origin and the area you would like to avoid, you can pre-allocate obstacles in the 2D map. That should solve the issue.

[rain1204]

Okay, I will try adding the bounding box.

I have another question. During the exploration process, I sometimes notice that the waypoints can get stuck, causing the robot to move back and forth between the same two nodes repeatedly. This issue seems to resolve only if I manually guide the robot away from these nodes, at which point the waypoints may update. However, this doesn't always work.

Have you encountered this issue before, and if so, do you have any advice on how to handle it?

[caoyuhong001]

In the launch file, you can set <param name="enable_save_mode" value="true"/＞. It will detect the back-and-forth behavior and navigate the robot to the nearest frontier.

[rain1204]

Ok I will try this parameter if the situation occurs again.

I recently encountered a couple of issues while testing the exploration :

1. AttributeError : ' Nonetype' object has no attribute 'data'

Occasionally, this issue occurs, causing the waypoint to stop updating entirely. When this happens, I need to restart the exploration process to recover.

2. The current node should be removed

This second issue doesn't seem to affect waypoint updates, but it consistently shows up in the terminal during runtime.

Have you encountered this issue before?

Thanks in advance! 😊

[caoyuhong001]

I guess it may be because your environment is too dense and you set a relatively large map resolution. So sometimes the robot or the start position is on the obstacle in the map.

[rain1204]

My testing environment is indoors, with many tables and chairs. I have tried using a smaller map resolution, but I still encounter the AttributeError: 'NoneType' object has no attribute 'data' error after the robot has been moving for a while.

[caoyuhong001]

You can comment https://github.com/marmotlab/ARiADNE-ROS-Planner/blob/53f79b9aa859f9166672541bc7b8eb2b71a7960a/src/scripts/node_manager.py#L114, which will stop removing unconnected nodes, to see whether the robot can complete the exploration. It likely will cause other issues though, but can confirm whether it is because the start node is located in the obstacles. If you can share the RViz screenshot and visualize the graph, it will be easier for me to evaluate the potential reason.