AI Racing Drone Could Transform Warfare — But Not Yet, Says Maker

A couple of years ago I wrote about the work of Davide Scaramuzza and his colleagues at the University of Zurich’s Robotics and Perception Group, and the development of a racing drone able to match human skills. Scaramuzza’s team went on to develop an AI system called Swift able to learn real flying skills in a virtual world. The latest breakthrough, published in Nature, has seen their deep-learning system beat human champions.

The world has moved on in the last couple of years: FPV racing drones have gone from being a niche e-sport only of interest to fans to a significant weapon in the Ukraine conflict with thousands of FPV kamikazes flown by both sides. Will Scaramuzza’s technology enable flying killer robots?

Racing Machines

The initial versions of the system used external motion trackers to follow the racing drones, giving them an unfair advantage over human pilots who were limited to the point of view of the drone’s camera. This setup also required sensors positioned around the racing arena which would not be applicable for most real-world situations. Swift is very different.

“After lot of work, we were recently able to outrace the world champions using only onboard sensors,” Scaramuzza told Forbes. “It’s based on an AI algorithm trained via reinforcement learning — trial and error, like the one of AlphaGo.”

AlphaGo developed by Google DeepMind was the first AI to defeat a human champion at the game of Go in 2015. It is still the strongest Go player in the world. Previous Go computer programs played at an amateur level, because the standard methods of testing all possible moves could not cope with the sheer number of possibilities. AlphaGo uses deep neural networks to get an overview of the Go board, then home in on possible strategies and evaluate each one.

Swift applies similar techniques to steering high-speed flight around a racing course, applying AI to a real-world contest.

“We are very excited as this is the first time that AI beats a human in a physical sport designed by and for humans,” says Scaramuzza. “Previous AI wins, like DeepBlue, AlphaGo, StarCraft, GT-Sophie, were limited to board games or simulation.”

Swift flew racing quadcopters against three humans: Alex Vanover, the 2019 Drone Racing League world champion, Thomas Bitmatta, twice MultiGP International Open World Cup champion, and Marvin Schaepper, three-time Swiss national champion, on a professional course with seven square gates.

The human pilots had a week to practice on the course before going head-to-head against Swift. Swift won several races against each of the human pilots and recorded the fastest track time. Scaramuzza says this is the first time that an autonomous mobile robot achieved world-champion-level performance in a real-world competitive sport.

“The AI is superhuman because it discovers and flies the best maneuvers, also it is consistent and precise, which humans are not,” says Scaramuzza. He notes that, as with AlphaGo, Swift was able to use moves — in this case flight trajectories — which the human champions did not even think were possible.

Taking AI to War

The implication is that AI-controlled FPVs can outperform humans. That has significant implications for warfare, where FPV drones with explosive warheads are tremendously effective human-guided weapons. A $400 FPV with the warhead from an RPG
RPG
rocket launcher can knock out a tank, personnel carrier or artillery piece from several miles away, or chase down and destroy a truck traveling at high speed. They are cheap enough to use against individual footsoldiers and can dive into trenches. But it requires a skilled human pilot. Ukrainian sources say the training takes around a month to achieve proficiency, and many people fail the course.

FPV success rates appear to vary wildly, with different sources citing 20%, 30%, 50% or 70% — much appears to depend on the exact situation, the presence of jamming, and the skill of the pilot. A super-skilled AI might push that rate far above 70%, negate any risk of jamming and enable fleets of smart FPV drones to attack simultaneously without human operators.

Slowing the Killer Robots

Scaramuzza notes that Swift could not easily be turned into a weapon.

“There are many open challenges that prevent the current system being deployed in warfare scenarios,” he says.

Swift relies on having reliable information on the speed, location and orientation of the drone in real time. This is far more challenging outdoors where there are changes of illumination, wind gusts and other variables to contend with.

Also, Swift has to learn the course ahead of time to work out its flight path.

“The current system only works for drone racing and for a specific racing track of which you perfectly know the map,” says Scaramuzza.

The neural network which navigates through the gates is trained specifically for that layout . The other problem is that Swift trains on a specific setup and if conditions change – for example the wind changes direction – all its learning may be wasted.

“Swift’s perception system and physics model assumes that the appearance of the environment and its physics are both consistent with what was observed during training,” says Scaramuzza. “If this assumption fails, the system can fail.”

This type of problem of drift, where the situation an AI has trained on does not quite correspond to reality, is well known. The U.S. Air Force is already developing a system to update its AI-controlled drones on the fly, retraining them using data gathered by the drones which have encountered problems. But nobody has yet looked at whether this type of approach could be applied to Swift, and Scaramuzza believes it will take a while.

“I think it will take decades before such a technology can be used in the military field,” says Scaramuzza.

This may be pessimistic, especially if large amounts of resource (and money) are assigned to the problem. But it does suggest there may be a breathing space before highly efficient AI-controlled killer drones swarm over the battlefield.

“This result should open our eyes on what AIs can potentially enable if unleashed,” says Scaramuzza. “That’s why in numerous talks at the United Nations, I always argue that we need a Geneva convention to stop killer robots.”

Years of UN discussions on this topic have so far produced little; the current proposal is a call for some sort of international law by 2026 but previous attempts to draft legislation have been watered down due to various states (including the U.S.) pushing back against binding agreements .

Swift shows what AI powered drones can do, and could lead to cheap Slaughterbot-type quadcopters being deployed on a massive scale. The defenders of Ukraine may think that is exactly what they need – and so might the invaders.