Yes, the IMU was only used to improve the accuracy of the encoder odometry.

Interesting, thanks for the info. When you’re talking about odometry providing the quadrant, was that the odometry + IMU? Or was IMU used for something else?

I modified Kris Winer’s implementation of Madgwick’s AHRS to do the 9 DoF sensor fusion. Since I was using ROS I used a filter from the robot_localization package. Most of the work was figuring out the proper covariances.

The point cloud data from the LIDAR is the important piece since I wasn’t technically navigating based on the odometry data in the traditional sense. In such a small space the LIDAR would know precisely where it was as long as it knew which of the four quadrants it was in, which is what was provided by the odometry. Since I knew that it would take significantly longer than the competition run time for there to be any significant error in the odometry I could get away without completing the loop.

Your localization block diagram shows wheel odometry and IMU data going into a Kalman filter, and you mentioned using a fusion algorithm for the IMU data. Was that AHRS fusion algorithm separate from the Kalman filter? Assuming it was, did you feed the position data from the Kalman filter back into the fusion algorithm? Or was the fusion algorithm only producing rates, which were used to update the position estimate in the Kalman filter?

I’m imagining that if the AHRS fusion and odometry algorithms both maintained an estimated pose without any feedback, they would diverge from each other and become less and less useful.

Hmm I might give the ODroid board a try — looking at their website their newest board has some pretty impressive specs!

I would recommend the Teensy boards over Arduino if you want to run ROS nodes on a microcontroller — they outperform the Arduino boards in practically every metric and most Arduino libraries will work on the Teensy. I almost always opt for using a microcontroller when dealing with encoders, high frequency control loops, or anything that has “realtime” requirements.

I’m going to do a post soon on running ROS nodes on a microcontroller and a SBC like a Pi — looks like a lot of people are interested in that. It’s really pretty straightforward and rather formulaic but it does require more than surface-level knowledge of C++.