Hi @dinesh-varun-kandiyappan ,
Please see follow up responses
If your IMX412 camera fails fisheye calibration, this is probably because the lens is not actually a fisheye. Please see the following post regarding a recent lens change : https://forum.modalai.com/topic/5321/imx412-m12-lens-difference
the new lens is rectilinear, so you should use the non-fisheye (plumb bob) model
The branch that saves the TOF calibration to a file is this one : https://gitlab.com/voxl-public/voxl-sdk/services/voxl-camera-server/-/tree/save-tof-v2-lens-cal . Sorry, we have not mainlined it yet. i will see if we can do that in the near future. see relevant post : https://forum.modalai.com/topic/5054/where-to-find-the-tof-and-hires-sensors-calibration
You may consider replacing your lens on IMX412 from the narrow one you have to a wider lens. We have the wide ones in stock and can get you a custom order if needed. Then you will see more features. The fisheye lens is about 30 degrees wider (horizontal FOV) than the non-fisheye.
VIO Features - maybe worth starting a separate post because it could be a long discussion and we could focus on that. Maybe make a new post with more details and we can get more folks on our team involved to look into it.
5.0 . PX4 params for Starling 2.. it looks like you may be correct.. double checking..
7.0 . We are testing Open Vins with rolling shutter cameras, let's discuss in another thread (see above).
8.0 . I don't think we have substantial updates to voxl-tflite-server.. will check.. and yes, qrb5165 is getting relatively old, so support for newer models can be difficult to get.
9.0 . If you are going to load up the cpu / gpu substantially, then a fan / heat spreader is needed. In general this is true for many mobile SoCs - they are powerful and efficient, but cannot sustain running at high utilization for long -- even phones can overheat and employ thermal throttling. My suggestion is to build up features / components one at a time and check your cpu / gpu utilization and thermal. Sometimes inefficient implementation of one component can overheat the whole system. also important to keep track of what resolution images you are using and make sure that you are not sending uncompressed images across regular (non-ION) pipes between processes..
before putting the second VOXL2 on the drone, you should proof everything on the bench. Starling 2 could certainly fit an extra VOXL2 mini. We don't have any official dates for VOXL3 yet, so it's best to focus on VOXL2. I will see if there is any info about VOXL3 we could share right now in terms of supporting newer inference models.
Can you please clarify what you mean by control policy? At what level are you training the control model (inputs and outputs?).
D
