Problems when running the examples

[duyanwei]

Hello, and thank you for open-sourcing this work—great job!

I was wondering if you have any recommended examples to start with. I’ve created a simple figure-8 trajectory with simulated IMU data (400Hz), poses(10Hz), and odometry(20hz) , and I’d like to try running your examples with it.

I began with pure_imu_integration and imu_pose3_fuser. However, pure_imu_integration doesn't seem to perform integration—the state remains at the initial pose. Meanwhile, imu_pose3_fuser crashes after a few iterations and throws exceptions from GTSAM. All the simulated data are with the lowest noise level.

Could you provide guidance on resolving these issues?

Additionally, it seems both examples require implementation to search for static transformations and parameters loading. I’ve hard-coded them for now, but does this indicate that these examples are still incomplete or not ready for direct use? Given that my goal is to fuse multiple pose or odometry sources with IMU data, which example would you recommend as a starting point?

Thanks again for your work and support!

[nubertj]

Hi @duyanwei,
Thanks a lot for your inquiry. These are exactly the two packages I haven't tested in a long time, just made sure they are still compiling.
However, in principle they should work.
for pure_imu_integration one has to still initailize the graph with inital position and yaw, otherwise it won't work. But for normal IMUs this is not very practical. We just used it for tests with a FOG IMU (super expensive one) where one can perform integration for dozens of seconds without too much drift.
For imu_pose3_fuser I haven't tested this recently, but I used it in the past to just fuse in IMU and pose which are expressed in the same sensor frame.

I can upload the bags from the paper in a few days to run the experiments for anymal and excavator. Meanwhile we could try the other two examples. Could you send me your data potentially? Otherwise I might be able to find some other data where IMU and pose are expressed in the same sensor frame.

[duyanwei]

Thank you for your reply. @nubertj

I've attached a ROS bag containing a simulated figure-8 trajectory for testing. The dataset includes ground truth odometry (100Hz), LiDAR poses at 10 Hz, visual poses at 30 Hz, and IMU data at 400 Hz. All data are already aligned in the same sensor frame, so no extrinsic calibration is required. I have added gaussian noise to the position components: 0.5 cm for the LiDAR poses and 2.5 cm for the visual poses.

It would be greatly appreciated if you have time to check whether the two examples can run successfully, especially with IMU data and multiple pose sources. I'm particularly interested in confirming that they handle these inputs correctly.