Questions Regarding Trajectory Tracking Error and Flight Testing in Gazebo

[Freq-q]

I trained the model using data collected from both the simplified simulation and Gazebo. However, during testing, I observe a trajectory tracking error in the range of several hundred millimeters, while the paper reports tracking errors of only a few tens of millimeters. I have carefully reviewed my simulation settings, data collection process, and training procedure, but I cannot identify any obvious discrepancies. Could this be due to simulation parameter mismatches, data quality issues, or other aspects of the testing process?

[Tim-Salzmann]

Hi Freq,

Thanks for your interest in our work and for reaching out to us.

Today, I tested the code in the GitHub repo by following the exact README instructions. Unfortunately, I was not able to reproduce your problems.

Specifically, I ran the simplified simulation (loop trajectory) without any model, which resulted in an RMSE of 241mm. With a trained model, trained on data from the simplified simulation, I got results of 20mm RMSE. This aligns with what we report in the paper (Table 3). Given that you consistently get error results in the magnitude of no model used, please verify that you are actually using the trained model in the controller (e.g., params in [1] as described in the README).

I took the opportunity to update the requirements to up-to-date versions. You could try to replicate my setup.

Let me know if this helps!

Thanks
Tim