In this project I want to test the performance of a Kalman filter. To achieve this, I built a robot controlled by the Raspberry Pi.
My plan was to create a drivable robot that would constantly track its position (using various sensors and a Kalman filter). In the end, the resulting track should be compared to a ground truth (for now, I only have manual verification).
For example, I often tried to drive something like an eight and then look at the track to compare its similarity to an eight.
The main workflow of how to use the robot is to drive around using a video game controller. Then in the end the calculated track is plotted and stored in a file so that someone can review it.
The robot is controllable via a video game controller that is connected with bluetooth. The east button of the controller can be used to stop the tracking.
I used the following sensors:
- GPS Sensor (position measurement)
- IMU Sensor (orientation and acceleration)
- Optical Flow Sensor (velocity)
For most of the sensors I needed to write my own driver since I was using rust (this was on purpose). Therefore, I believe there are some bugs in my code. Although, I was quite happy with the measurements.
In most of my runs, I started somewhere, then drove around a table and back to where I started. So the resulting shape should have some similarity to a balloon. Below are some example tracks I got on my runs.
Example of a good track:
Example of a bad track:
As you can see in the good track I have the desired balloon-shape (quite accurately). But in the bad track I have more of a U-shape.
So there seems to be some offset between the way to the table and the way back from the table. And I had this U shape quite a lot. It seems that the GPS sensor has some problems depending on the with the direction I was going.
But I am still quite happy with the results. In particular, the Kalman filter implementation worked quite well. For example, I once set the optical flow error to 0, and then I could see the balloon shape as well as the drift.
I would be interested in testing other sensors in the future. For example, I would be very interested to see how a LiDAR sensor would affect the system. Especially since the current system relies quite heavily on the GPS sensor, it would be interesting to see what effect that would have.
Frederik Böcker