Interesting looking at the odds analysis. From the Binomial distribution at odds of hitting of 0.75, you'd need to attack with 8 drones to have a 90% chance of one getting through. There's going to be an arms race in terms of numbers of attackers Vs defenders in the future.
It wouldn’t surprise me. If you look at how the US Navy defended itself against Japanese kamikaze planes during WWII it was with a wall of exploding lead.
Interesting, visionary playing main role where the AI is concerned, we designed 24/7 low light sort to long range vision system for more details https//nightpulser.com
Nice research and truly unique as its trying to quantify possible results.
One think to consider is that AI, although now is used mainly for machine vision, i.e identification, it is probable that will be able to estimate the possible flight paths based on the drone characteristics and thus increase the chances of a hit.
Also expect to see multi-modal identification and tracking capabilities using passive detection of RF/thermal signatures in addition to visual.
Classic methods for tracking and flight path estimation are faster and more reliable and likely to stay that way -- the hard vision part is getting accurate, consistent object detections over time from which to form a trajectory for predictive modeling.
Smarter AI might make a difference for more contextual flight path prediction, like dealing with evasive flight patterns, but IMO there's not room there either. If I'm programming an FPV flight path and I'm worried about something like a Bullfrog, I'm probably going to add random jitter in some manner that spoils trajectory prediction by any means. And meanwhile, your fancy AI has a bigger compute requirement (more weight, cost, and points of failure) and higher latency that's likely not compatible with the sort of tracking needed to down a drone at 800m with a machine gun.
Then, it's a cat-and-mouse game between how responsive your M240 turret can be, how many rounds your turret can get in the air, how much energy you can afford to put into your drone's evasive fight pattern, and how subtle your jitter can be (smaller, faster displacements are less predictable but may not move your drone enough to dodge the volume of rounds coming its way).
There will likely also be a cat-and-mouse game of trying to fool the visual identification algorithms by altering drone profiles (computer vision algorithms trained to detect e.g. drones can be very brittle if they start seeing shapes they weren't trained on) and controlling the visual landscape of the approach (come in low in front of cluttered terrain). There may be an opportunity for a smarter, slower AI system to provide context clues to a low-latency embedded vision system to dynamically alter its recognition capabilities to adapt to unexpected changes in the field.
Agreed about multi-modal sensors likely to play a larger role as the core capabilities are proved out. I imagine the full range of relevant passive sensors will be explored.
Interesting looking at the odds analysis. From the Binomial distribution at odds of hitting of 0.75, you'd need to attack with 8 drones to have a 90% chance of one getting through. There's going to be an arms race in terms of numbers of attackers Vs defenders in the future.
Think they’ll bring back specialized flak?
It wouldn’t surprise me. If you look at how the US Navy defended itself against Japanese kamikaze planes during WWII it was with a wall of exploding lead.
7.62 snakeshot case closed
Interesting, visionary playing main role where the AI is concerned, we designed 24/7 low light sort to long range vision system for more details https//nightpulser.com
Bravo.
Has anyone else done this yet?
That’s always instructive.
Nice research and truly unique as its trying to quantify possible results.
One think to consider is that AI, although now is used mainly for machine vision, i.e identification, it is probable that will be able to estimate the possible flight paths based on the drone characteristics and thus increase the chances of a hit.
Also expect to see multi-modal identification and tracking capabilities using passive detection of RF/thermal signatures in addition to visual.
Classic methods for tracking and flight path estimation are faster and more reliable and likely to stay that way -- the hard vision part is getting accurate, consistent object detections over time from which to form a trajectory for predictive modeling.
Smarter AI might make a difference for more contextual flight path prediction, like dealing with evasive flight patterns, but IMO there's not room there either. If I'm programming an FPV flight path and I'm worried about something like a Bullfrog, I'm probably going to add random jitter in some manner that spoils trajectory prediction by any means. And meanwhile, your fancy AI has a bigger compute requirement (more weight, cost, and points of failure) and higher latency that's likely not compatible with the sort of tracking needed to down a drone at 800m with a machine gun.
Then, it's a cat-and-mouse game between how responsive your M240 turret can be, how many rounds your turret can get in the air, how much energy you can afford to put into your drone's evasive fight pattern, and how subtle your jitter can be (smaller, faster displacements are less predictable but may not move your drone enough to dodge the volume of rounds coming its way).
There will likely also be a cat-and-mouse game of trying to fool the visual identification algorithms by altering drone profiles (computer vision algorithms trained to detect e.g. drones can be very brittle if they start seeing shapes they weren't trained on) and controlling the visual landscape of the approach (come in low in front of cluttered terrain). There may be an opportunity for a smarter, slower AI system to provide context clues to a low-latency embedded vision system to dynamically alter its recognition capabilities to adapt to unexpected changes in the field.
Agreed about multi-modal sensors likely to play a larger role as the core capabilities are proved out. I imagine the full range of relevant passive sensors will be explored.