PX4 Vision Drone Unboxing and Testing - Best drone for dev yet?

PX4 Vision unboxing with professional reviews
PX4 Vision obstacle avoidance test in different environments and modes
Path planning and obstacle avoidance test in Mission Mode
music by audionautix.com


Hey @MasterPa thanks for posting, really cool video, looking forward to part 2.

Good demonstration video, which illustrates some problems with obstacle detection. I’m really surprised that the manufacture used the Structure.io depth sensor rather than the superior (and cheaper) Intel Realsense 435s. Anybody know why that is?

The realsense had some serious issues with auto-exposure for a long time, which made it basically unusable in bright sunlight. They have been fixed now as far as I understand, but were still an issue during the design phase of PX4 Vision. Also the Structure Core camera has a significantly longer range (~12m vs ~ 8m for the realsense) and the pointcloud quality is superior.


Are there plans to make this drone platform commercially available? Is there a parts list anywhere I can reference to get started?

@baumanta, according to official spec

Depth Min/Max
Z Distance 0.3-5m. Up to 10m (Depending on scene & lighting conditions.)
Depth FOV (H x V x D) 59° x 46° x 70°

Maximum Range:
Approx. 10 meters. Accuracy varies depending on calibration, scene, and lighting condition.
Depth Field of View (FOV):
87°±3° x 58°±1° x 95°±3°

BTW Has anyone tried new ZED?

BTW For Structure Core I don’t see any official ROS package? Seems like they are more oriented towards iOS, VR/AR, etc. not robotics

I’ve got both the Structure Core and the Realsense 435 and in practice they have similar range in sunlight (around 8m). Aside from being cheaper, the Realsense requires much less compute power (since it does most of its work with its onboard Myriad processor) and the SDK works much better on Linux (Occipital never released its SLAM library on Linux, sadly). Both of them have global shutters, which is important for motion. Also the Realsense 435 works well with the Realsense T275 visual odemetry sensor, which make it easy to add VIO to the drone

Yes, that’s correct. They started with a focus on iOS and haven’t really moved far from that

Sadly, the ZED2 has a rolling shutter (not global), which is inappropriate for fast moving vehicles like drones. It also costs $450!

The ZED camera requires an Nvidia GPU to run the SKD on, which is another reason (additional to the ones mentions by @Chris_Anderson) why we do not use it.

also we are talking about the Structure Core camera, not the Structure Sensor, these are two different models. There is an opensource ROS wrapper for it (but not on git) you can download it from Occipitals website.

@Chris_Anderson We have definitely seen more than 8m outdoor when setting the camera to long range mode. Also we connect the camera over USB 2 on the vehicle, which causes the quality of the realsense data to drop drastically. On the Structure Core it does not seem to have much influence.

1 Like

Yes, you definitely have to connect the Realsense to a USB3 port. The Up board should have those

If the weather improves here, I’ll show the outdoors performance of the Structure Core and Realsense 435 side-by-side

The problem is not the availability of the USB3 ports, but that USB3 jams the GPS. We have not found a way yet to shield against USB3 reliably so we are not using any USB3 cables anywhere on our drones.
I’m really looking forward to the comparison, please let me know what your findings are.

1 Like

Interesting! I didn’t know USB3 interferes with GPS.

I did a quick side-by-side test of the Realsense vs the Structure core on a somewhat overcast day. You need to turn the Structure Core gains and aperture/shutter speed settings down all the way to get anything outdoors. Basically, in that lighting the Structure Core has a range of 10-12m while the Realsense is 8m, but the Realsense has a RGB color camera and higher-resolution image:

1 Like

The Intel Realsense 435 can now capture at 300FPS for fast-moving vehicles:


Wow, this looks fantastic.

Is there any relevant parameters about the obtacle avoidance?

Nice sharing. So where is the part 2 link?