VIO drifting when using TFlite Server

[Ege Yüceel]

Hi ModalAI team

I’m working with a Starling 2 drone running SDK 1.4.0. In a recent project, I integrated a TFLite server to command the drone in certain directions using VIO (qvio).

Here’s the situation:

In Position mode, VIO works fine. It can handle mild to moderately agile maneuvers with good stability. However, once I start using the TFLite server to command the drone (via MAVSDK Python Docker, listening to TFLite outputs from a mounted pipe), VIO becomes unstable. The issues I observe: During slightly rapid motions, I get errors like “accelerometer fail” or “height estimate error”. Over time, VIO itself fails completely and I have to restart.

I am presuming this can happen because of overheating. When the neural network runs, the VOXL2 temperature often reaches 80–85°C, sometimes even above 90°C when I dont start the propellers quickly. This could:

-Directly harm the VOXL2 hardware.

-Disrupt IMU calibration at high temperatures.

I’m thinking about adding a cooling fan or redoing the IMU calibration (hot calibration in particular).

What would be your suggestions or other potential reasons for this case?

Thanks!