AI Quadcopter 'Swift' Beats Top Human Drone Racers (gizmodo.com) 19
An autonomous, artificial-intelligence-powered drone called Swift has beaten humanity's best drone racers. "The AI-equipped drone, developed by researchers at the University of Zurich, came out on top in 15 out of 25 races and recorded the single fastest lap time," reports Gizmodo. The findings have been published in the journal Nature. From the report: Swift beat the humans in the niche but growing sport of first-person view drone racing. Human competitors navigate using a headset connected to a camera on their drones to pilot a quadcopter through complex obstacle courses at extreme speeds, with the goal of finishing the race with the fastest time and avoiding taking too much damage in the process. Drones in these races can top 50 miles per hour when they're really buzzing. The [video here] shows Swift battling it out against the human-controlled drones.
Swift emerged victorious in 15 out of the 25 total head-to-head races against the human pilots and clocked the fastest overall lap time at 17.47 seconds. That brisk lap time was nearly half a second better than the best human. The three human competitors, Alex Vanover, Thomas Bitmatta, and Marvin Schaepper, have each won drone racing championships in the past. In this case, the human competitors had a week to learn the new course and train for the race. During that same time, Swift was training as well but in a digitally simulated environment meant to resemble the course. Swift, according to the paper, used deep reinforcement learning while in the simulation along with additional data collected from the outside world.
During the actual race, Swift would take in video collected by its camera and send that to a neural network capable of identifying the gates it had to fly through. A combination of onboard sensors are then used to aid the drone with positioning, speed, and orientation. All of this happened autonomously, at extreme speeds. The researchers noticed some interesting differences in the ways Swift approached the course as opposed to its human competitors. The autonomous system, they noted, was more consistent across laps and appeared to take tighter turns. Those tight turns can add up and give a drone an edge in a race by repeatedly shaving off fractions of a second from lap times.
Swift emerged victorious in 15 out of the 25 total head-to-head races against the human pilots and clocked the fastest overall lap time at 17.47 seconds. That brisk lap time was nearly half a second better than the best human. The three human competitors, Alex Vanover, Thomas Bitmatta, and Marvin Schaepper, have each won drone racing championships in the past. In this case, the human competitors had a week to learn the new course and train for the race. During that same time, Swift was training as well but in a digitally simulated environment meant to resemble the course. Swift, according to the paper, used deep reinforcement learning while in the simulation along with additional data collected from the outside world.
During the actual race, Swift would take in video collected by its camera and send that to a neural network capable of identifying the gates it had to fly through. A combination of onboard sensors are then used to aid the drone with positioning, speed, and orientation. All of this happened autonomously, at extreme speeds. The researchers noticed some interesting differences in the ways Swift approached the course as opposed to its human competitors. The autonomous system, they noted, was more consistent across laps and appeared to take tighter turns. Those tight turns can add up and give a drone an edge in a race by repeatedly shaving off fractions of a second from lap times.
Interesting (Score:3)
If I read correctly this is entirely ONBOARD AI, so it doesn't rely on external computing power. If this eliminates communication these could have real implications on range, jamming, etc.
Re: (Score:3)
It only beats humans until you move a gate a few inches.
It's nowhere near as impressive as the folks selling stock in the company would have you believe.
(It is impressive, just not to the degree claimed.)
Re:Interesting (Score:4, Insightful)
OK... now give it a terrain map for general guidance and let it use its onboard smarts for low-altitude obstacle avoidance while staying close to the ideal course.
Now you have a smart suicide drone that can fly through cities or forests and get to within feet of the intended target, all while more or less impossible to shoot down.
It's not a racing system, it's a missile guidance system.
Re:Interesting (Score:5, Informative)
NPR did a piece on it this morning. One of the things they mentioned was that the "training" phase involved it going through the course several hundred times over the course of hours. And even then, if it got bumped by another drone, it, in their words, spun out and crashed because it didn't know what to do.
It doesn't operate from terrain maps, it operates from running into obstacles in the course over and over until it misses them all.
You want to shoot it down? Walk a few feet to the side.
Re: (Score:2)
Insightful and nicely put :)
News at 11 (Score:3)
Computer does something better than humans at something computers do better than humans.
Re:News at 11 (Score:4, Insightful)
Computer does something better than humans at something computers do better than humans.
Shocking? Not really.
Important? Very.
Right now drone warfare is a huge factor in the war in Ukraine. This includes "short range" front line drones, used to attack personnel and equipment. And long range drones taking the place of missiles.
For the most part these drones are either remotely controlled, or given fairly rudimentary self-piloting capabilities (ie, fly to GPS coordinate). In both cases they're susceptible to electronic war countermeasures, either jamming (for the remote control) or GPS interference.
Now, replace that with a drone that is fully antonymous. Over short range erratically skimming along the ground (like one of these quad-copters) until it sees an enemy to slam into and destroy. Over the long range moving along more slowly, indistinguishable from a bird, navigating using landmarks and the position of the sun and stars until it finds its target.
Those are REALLY hard threats to stop.
So yeah, these kind of advancements are important.
Re: (Score:2)
It memorized a path and then replayed it faster than a human.
Ooooh.
Re: (Score:2)
It memorized a path and then replayed it faster than a human.
That's not a thing that happens. You can't just record moves and then play them back and win the race. It had to do the processing in realtime, as a human would, and that's what it did.
The humans also trained on the course, while the neural net was trained in software, so the humans had the advantage of training on the real thing.
Re: (Score:2)
It memorized a path and then replayed it faster than a human.
That's not a thing that happens. You can't just record moves and then play them back and win the race. It had to do the processing in realtime, as a human would, and that's what it did.
The humans also trained on the course, while the neural net was trained in software, so the humans had the advantage of training on the real thing.
This is true. However, it's like validating with the same training data, which is not very impressive and would heavily favor a machine. What would have been more impressive is validating both the humans and drones on a new course that had similar features.
Unfair comparison (Score:4, Insightful)
Re: (Score:2)
It doesn't matter whether the neural network is hosted on a machine inside of the drone or not. You could send the data out to be processed just like a human does it, or not, depending on whether you can fit enough hardware on the drone to do the job.
I didn't read enough of the paper to determine whether the neural net is getting feedback from the IMU or it's all done visually. If it is, then THAT is an unfair advantage. If the humans were actually riding the drones, it might give THEM an advantage over an
Re: (Score:2)
From the video, it mostly uses inertial measurement, with visual recognition of the gates to correct for drift.
Re: (Score:2)
Yeah, that's a definite advantage. It's like the difference between driving, and a driving simulator.
Changes (Score:2)
Now change the track layout and see how long it takes to adapt. I'm guessing the human pilots don't take as long.
Promising (Score:2)
Re: (Score:3)
GPS isn't nearly fast enough for the kind of maneuvers in FPV races. Ignoring that most courses are indoors, GPS satellites transmit ephemeris data at ~50bps. A good GPS receiver gives smoothed position estimates from that at 5hz. A fifth of a second isn't terrible for position updates, but it would be a lifetime in drone racing. A competition legal racing drone can move 4-5 meters or completely reverse its direction of travel in that time.
For perspective on how fast these are, it's worth watching this
AI drone race disaster of 2033 (Score:3)
It was a special day at the Drone Arena, the guest of honor was none other than little Billy. A child who overcame poverty and physical disability to write the New York Times best selling novel at age 9. Unfortunately for Billy the AI was never trained to avoid spectators on the field in wheel chairs and his torso was the shortest path to the finish line for the drones.