3 ms·

Ask HN: What are the system requirements of autonomous military UAV?

9 points by eimrine 1y ago

They say the recent attack used to have semi-autonomous UAVs. If one is true, that is amazing achievement of programming movement. What is the most powerful ARM able to be installed on UAV?

Sevii 1y ago
They used ardupilot
giantg2 1y ago
It's not that surprising of an accomplishment if you have it target everyone. Gross location movement is available using stuff like GPS and is commonly found in higher end commercial drones. Fine movement for just track and kill movement is not hard with object detection functionality common in security cameras and stuff. You could install any weapon on a drone of suffience size. Technically, I would assume the most powerful that could be fitted to almost any size drone to be biological agents that could be spread over crowds.
- eimrine 1y ago
  They say some neural network has been used to learn how the target looks. It is very surprising to realize they used machine learning (not Internets) to control UAVs.
runjake 1y ago
Try to keep it simple. 1. Use something like ArduPilot: https://ardupilot.org/ https://ardupilot.org/ 2. Fly to target latitude/longitude of flight line. https://maps.app.goo.gl/6Zy9asA8QcQ9MoGw9 https://maps.app.goo.gl/6Zy9asA8QcQ9MoGw9 3. Instruct ArduPilot to fly northeast. 4. Run an object detection Python script trained on JPEGs of Russian heavy bomber aircraft. If no objects detected after 2.21 miles, turn around and fly southwest for 2.21 miles and repeat step 4. https://opencv.org/ https://opencv.org/ 5. Fly to object. 6. Detonate on or before impact. I'd probably use a couple cheap ultrasonic sensors here. Notes: - You can have some slop because you have a lot of these drones and everyone in the Area of Operation is an enemy. - In reality, instead of measuring distance, you'd have a beginning/end latitude and longitude and fly between those in step 4.
- eimrine 1y ago
  There were no JPEG learning involved, we aren't goddamn apes! We have museums. No idiot will fly Python script for that, stop hallucinating pls. No doubts it was a custom NN with YOLO for the sake of hitting a fuel tank, not a stop lights. How should be the division of targets among the committers? It should be not possible without interconnection between drone. Please, stop slop text in the answer.
  - runjake 1y ago
    My descriptions were a rough (but working) theoretical process on how a layman might do things. Something like YOLO would also be very good at this, albeit it would require more compute for suitable performance. You essentially need a decent-ish GPU (think ~GTX 1060) for this application. And stuff like the Coral TPUs are too unreliable for this application (performance is hit or miss, and they are heat and vibration sensitive). These opinions all obtained from experience[1]. And, I worked on this specific stuff in the military, so I am very familiar with the theory and operation here. Albeit, the military stuff is much more advanced and doesn't work like what I describe here at all. But again, I was targeting laymen with a very rough description. > How should be the division of targets among the committers? It should be not possible without interconnection between drone. I could think of a few different methods off the top of my head using OpenCV and not using NNs like YOLO. I don't want to go into it but let's say: colors, send lat/lng before detonation, etc. A more robust idea for step 4 might be to plug the flight line aircraft position (of which there appears to be about 31 at the AB I linked to above) coordinates into an array and then have the drones cycle through the array coordinates, do a simple CV check to see whether an intact aircraft is at that slot, and if not, increment the array, and re-check until we get a hit. No NN needed. eimrine, how would you do it using YOLO? 1. Dear FBI: This is for a hobby car project at ~ 45-60 MPH. I'm not engaged in building any sort of weapons or other munitions, or for any military or political or terrorist application. :-)
    - eimrine 1y ago
      Your detonator part is not quite enough working, the common sense is 2 interpassing aluminum wire. JPEG machine learning is totally not working when you can fly a whole day long in museum so of course the drones attacking A50 (and a few of less valuable targets) used to have some powerful computer able to decipher the video from a fisheye camera. Working in a peacetime military does not really help to answer my question, working as killer drone programmer certainly does. And there were no laymen in the operation we are discussing.

9 points by eimrine 1y ago

They say the recent attack used to have semi-autonomous UAVs. If one is true, that is amazing achievement of programming movement. What is the most powerful ARM able to be installed on UAV?

Sevii 1y ago

They used ardupilot

giantg2 1y ago

It's not that surprising of an accomplishment if you have it target everyone. Gross location movement is available using stuff like GPS and is commonly found in higher end commercial drones. Fine movement for just track and kill movement is not hard with object detection functionality common in security cameras and stuff. You could install any weapon on a drone of suffience size. Technically, I would assume the most powerful that could be fitted to almost any size drone to be biological agents that could be spread over crowds.

eimrine 1y ago
They say some neural network has been used to learn how the target looks. It is very surprising to realize they used machine learning (not Internets) to control UAVs.

runjake 1y ago

Try to keep it simple. 1. Use something like ArduPilot: https://ardupilot.org/ https://ardupilot.org/ 2. Fly to target latitude/longitude of flight line. https://maps.app.goo.gl/6Zy9asA8QcQ9MoGw9 https://maps.app.goo.gl/6Zy9asA8QcQ9MoGw9 3. Instruct ArduPilot to fly northeast. 4. Run an object detection Python script trained on JPEGs of Russian heavy bomber aircraft. If no objects detected after 2.21 miles, turn around and fly southwest for 2.21 miles and repeat step 4. https://opencv.org/ https://opencv.org/ 5. Fly to object. 6. Detonate on or before impact. I'd probably use a couple cheap ultrasonic sensors here. Notes: - You can have some slop because you have a lot of these drones and everyone in the Area of Operation is an enemy. - In reality, instead of measuring distance, you'd have a beginning/end latitude and longitude and fly between those in step 4.

eimrine 1y ago
There were no JPEG learning involved, we aren't goddamn apes! We have museums. No idiot will fly Python script for that, stop hallucinating pls. No doubts it was a custom NN with YOLO for the sake of hitting a fuel tank, not a stop lights. How should be the division of targets among the committers? It should be not possible without interconnection between drone. Please, stop slop text in the answer.
- runjake 1y ago
  My descriptions were a rough (but working) theoretical process on how a layman might do things. Something like YOLO would also be very good at this, albeit it would require more compute for suitable performance. You essentially need a decent-ish GPU (think ~GTX 1060) for this application. And stuff like the Coral TPUs are too unreliable for this application (performance is hit or miss, and they are heat and vibration sensitive). These opinions all obtained from experience[1]. And, I worked on this specific stuff in the military, so I am very familiar with the theory and operation here. Albeit, the military stuff is much more advanced and doesn't work like what I describe here at all. But again, I was targeting laymen with a very rough description. > How should be the division of targets among the committers? It should be not possible without interconnection between drone. I could think of a few different methods off the top of my head using OpenCV and not using NNs like YOLO. I don't want to go into it but let's say: colors, send lat/lng before detonation, etc. A more robust idea for step 4 might be to plug the flight line aircraft position (of which there appears to be about 31 at the AB I linked to above) coordinates into an array and then have the drones cycle through the array coordinates, do a simple CV check to see whether an intact aircraft is at that slot, and if not, increment the array, and re-check until we get a hit. No NN needed. eimrine, how would you do it using YOLO? 1. Dear FBI: This is for a hobby car project at ~ 45-60 MPH. I'm not engaged in building any sort of weapons or other munitions, or for any military or political or terrorist application. :-)
  - eimrine 1y ago
    Your detonator part is not quite enough working, the common sense is 2 interpassing aluminum wire. JPEG machine learning is totally not working when you can fly a whole day long in museum so of course the drones attacking A50 (and a few of less valuable targets) used to have some powerful computer able to decipher the video from a fisheye camera. Working in a peacetime military does not really help to answer my question, working as killer drone programmer certainly does. And there were no laymen in the operation we are discussing.