@Maxwell-Schaefer Thank you! I was able to set the model match to off using the controller screen.

@tom Thank you. I will submit an RMA request.

Hi @mark ,

Thank you for the update. We will keep investigating the underlying issue that was a result of that questionable camera.

Meanwhile, if you think that the camera is defective and has no damage, you can fill out an RMA request and send it back for an exchange. I cant really comment on what could be causing this on the camera itself (it would be interesting to try it), but the camera is not serviceable by the user.

https://www.modalai.com/pages/rma

Alex

@jeremyrbrown5 we figured it out, we just changed the name of the model to yolov8n_float16.tflite. Pretty spot on now!

@Hunter-Scott Can you update to SDK 1.6.0? Then do the ver all and qshell voxl_esc status commands again and attach the output?

@taiwohazeez Can you post a video of it trying to fly?

Hi @Rowan-Dempster ,

Please take a look at this example (you can build and run it too) : https://gitlab.com/voxl-public/voxl-sdk/utilities/voxl-mpa-tools/-/blob/add-new-image-tools/tools/voxl-image-repub.cpp

This app can accept a regular image (RAW8 or YUV) and either re-publish it unchanged or crop and publish the result. Sometimes this is useful for quickly cropping an image that is fed into a different application that expects a smaller image or different aspect ratio.

The app shows how to subscribe and handle ion buffer streams.

Usage:

voxl2:/$ voxl-image-repub ERROR: Pipe name not specified Re-publish cropped camera frames (RAW8 or YUV) Options are: -x, --crop-offset-x crop offset in horizontal dimension -y, --crop-offset-y crop offset in vertical dimension -w, --crop-size-x crop size in horizontal dimension (width) -h, --crop-size-y crop size in vertical dimension (height) -o, --output-name output pipe name -u, --usage print this help message The cropped image will be centered if the crop offsets are not provided. typical usage: /# voxl-image-repub tracking --crop-size-x 256 --crop-size-y 256 /# voxl-image-repub tracking --crop-size-x 256 --crop-size-y 256 --crop-offset-x 128 --crop-offset-y 128

example re-publishing ion buffer image as regular image (which you can view in voxl-portal ) :

voxl-image-repub tracking_front_misp_norm_ion -o test

(you can see which ion pipes are available by running voxl-list-pipes | grep _ion)

Please note that without the previous fix that i posted above, the client process that receives and uncached ION buffer will incur extra CPU load while accessing this buffer. For example, the same voxl-image-repub client uses 1.7% cpu while republishing the normalized image (cached ion buffer), while using 7.3% cpu republishing an image from an uncached ION buffer. (cpu usage % using one of the smaller cores).

Please try and let me know if you have any questions.

I know this cached / uncached buffering may be a bit confusing, but i will document this a bit more to help explain it a little better.

Alex

@tom Thank you! Submitted the RMA today (22-NOV-2025). Did not get any number or anything, but the following message: a25e2aa8-af17-4b82-8b20-c78ef379e805-image.png

@Muqing-Cao Here you go: https://gitlab.com/voxl-public/voxl-sdk/services/voxl-vision-hub/-/blob/dev/src/vio_manager.c?ref_type=heads#L450

@Jakub-Tvrz said in Starling 2 Max for indoor:

We haven't bought anything yet, but we are deciding what will be the best for our warehouse inventory application. Of course, we considered the ToF option. But you still say: "however, it is not... great" - does that mean that Starling 2 is better for our application? Or does Dual vs Triple camera for Visual Inertial Odometry also have an impact?

Reply

Depends fully on your use case. I will say that PX4's instance of obstacle avoidance is not fantastic, but it is better than nothing. In my opinion I would go w/ the model that has all the bells and whistles in case you need the hardware and decide to develop your own collision avoidance or somethign of that nature.

@lhermosillo
That is harsh. Do you have a bare board, or one with some additions? If you can use the bare voxl2, try that. If you have been using the bare board all along, I'm stumped here. I'd maybe try to reinstall the modalai-qdl env, or use another image. Maybe the qdl command itself instead of the webUI. Hope you'll be able to save your board!

@aspen The landing detector in PX4 isn't perfect (obviously) so in this case it didn't detect the landing. I would map one of the switches to toggle between arm and disarm so that when the drone lands you can disarm via switch. Did you try the v1.16 package?

@N-Chen
To anyone reading this, maybe encountering the same issue in the future.
I've tested the board with a guy that has a lot of experience with the voxl2 board,
According to him, the board was faulty, probably some electrical issue, the amperage did not cross 0.1, and the board usually runs on something like 0.5.
I've replaced my board as well as my power-supply cables (which were also under par).

@Rowan-Dempster , you should use a monochrome stream (_grey), since QVIO needs a RAW8 image.

If you are not using MISP on hires cameras, that is fine, you can start off using the output of the ISP.

You should calibrate the camera using whatever resolution you decide to try. This is to avoid any confusion, since if you using ISP pipeline, the camera pipeline may select a higher resolution and downscale + crop. So whenever you are changing resolutions, it is always good to do a quick camera calibration to confirm the camera parameters.

When using MISP, we have more control over which camera mode is selected, because MISP gets the RAW data, not processed by the ISP, so we know the exact dimensions of the image sent from camera.

Alex

@joseph-vale ,

You did not provide the actual error that you are seeing, however I could try to guess what it is (even if not, the details below should probably help you anyway). The default build of voxl-opencv package does not have python3 support. So if you are using gstreamer with opencv in python and that is the error, you should install the voxl-opencv package that I built with python3 support

https://storage.googleapis.com/modalai_public/temp/voxl-opencv_4.5.5-3_arm64.deb source : https://gitlab.com/voxl-public/voxl-sdk/third-party/voxl-opencv/-/tree/add-python3-bindings

Below, i will assume that you want to grab images from Boson part of the Hadron, however similar approach should apply to the RGB camera in Hadron.

First you should test ability to grab images without python. You may need to replace your camera # depending on what camera id your Boson is.

IMPORTANT: make sure that voxl-camear-server is not running while you are trying to use gstreamer.

systemctl stop voxl-camera-server

Tip: you can actually stream video using X forwarding with ssh. This should stream live Boson feed from voxl2 to your linux machine:

ssh -Y username@<voxl-ip> gst-launch-1.0 qtiqmmfsrc camera=0 ! "video/x-raw,width=640,height=512,framerate=30/1" ! videoconvert ! autovideosink

display image directly in terminal as ascii:

gst-launch-1.0 qtiqmmfsrc camera=0 ! "video/x-raw,width=640,height=512,framerate=30/1" ! autovideoconvert ! aasink

and then finally, a python script that grabs h264 video from qtiqmmfsrc, decodes it and returns frames to python:

import time import cv2 #get RGB (BGR?) directly #stream='gst-launch-1.0 qtiqmmfsrc camera=0 ! video/x-raw, width=640,height=512,framerate=30/1 ! autovideoconvert ! appsink' #get h264 -> decode -> BGR stream='gst-launch-1.0 qtiqmmfsrc camera=0 ! video/x-h264,format=NV12,profile=high,width=640,height=512,framerate=30/1 ! h264parse ! qtivdec ! qtivtransform ! video/x-raw,format=BGR,width=640,height=512 ! autovideoconvert ! appsink' print(stream) vcap = cv2.VideoCapture(stream,cv2.CAP_GSTREAMER) frame_cntr = 0 while(1): ret, frame = vcap.read() if ret: frame_cntr += 1 print('got frame %d with dims ' % frame_cntr, frame.shape)

Hopefully, that works for you.

Final recommendation - if you use qtiqmmfsrc this way, the Boson data is processed in the Qualcomm ISP and unless your have a special tuning file for Boson, the processed output will have degraded quality. Boson, by default, outputs post AGC 8-bit image which is already processed and does not need to be further processed by the ISP. I am not sure whether you can get RAW8 data from qtiqmmfsrc (unmodified data from Boson).

We handle the above issue in voxl-camera-server by working with the RAW8 directly. We also recently started experimenting with 14bit pre-AGC data from Boson, which would need some processing before it is usable (if you are interested in that, i can share some more information).

Finally, if you would like to use voxl-camera-server, which is what we recommend and support, there is also a way to get encoded h264/h265 data into python (using our experimental pympa (python MPA bindings)). That is a topic for a discussion in another post, if you are interested..

Alex

@cbay ,

We support any camera that is compatible with UVC (plugged into a usb port). Please look into voxl-uvc-server. https://docs.modalai.com/voxl-uvc-server/

Regarding NIR, have you considered using a regular camera (most cameras are sensitive to NIR) and install a lens with a NIR pass filter (if you dont want to see visible light)? Then you could use one of our hi-res cameras like IMX412 or IMX664. I can get you spectral response plot of those sensors if you need.

Alex

@kabumei
Hello, my name is Naoki, I am keen on this project, could you send me a detail to major.gits501@gmail.com?

@Seungjae-Baek , the resolution of the camera should depend on the use case. However, it is also important to keep in mind what exactly you are doing with the images coming from the hires cameras. For example, if you set the resolution to 4K and you try to view those uncompressed images using voxl-portal, this process will be very taxing on the CPU for the following reason: when you view uncompressed images in voxl-portal, these images are encoded with a software JPG encoder and then sent from voxl2 to your browser. This process is very cpu-heavy. On the contrary, if you use h264 / h265 stream, it should be perfectly fine to encode 4K30 video to disk or even stream, since H264 / H265 encoding is done by a hardware encoder.

voxl-portal, actually does support showing h264 (but not h265) 30FPS streams, so that would be a lot more efficient for the CPU, since there would be no jpg encoding. Otherwise, if you are using raw frames (not _encoded) in voxl-portal , please keep in mind that you will always have a lot of cpu overhead. voxl-portal is designed for debugging / development purposes, so it's not necessarily the most efficient solution for video streaming. For real video streaming use cases, you would use h264 or h265 encoding and save to disk on voxl2 + stream the encoded video for remote viewing. You could encode the same camera source with two different resolutions / codecs / bitrates.

If you need help setting up a specific use case, please provide some details and I can help you further.

Alex

@Zachary-Lowell-0

Thank you for the feedback! This was very helpful

@박지현 Unfortunately, for Ardupilot, our ESCs are only supported with UART interface when using the VOXL 2 / VOXL 2 mini as the flight controller. When using Flight Core v2 that is not supported.