Fully autonomous drone outperforms two 'world-class' human drone racers

Cal Jeffrey

Posts: 3,119   +862
Staff member
The bleeding edge: Researchers at the University of Zurich (UZH) developed a machine-learning algorithm for controlling a quadcopter drone that can outperform professional drone racing pilots. The algorithm calculates "time-optimal trajectories" while also factoring in the drone's limitations.

The feat seems obvious at first glance—a machine learning system beat a human again, so what? However, pro drone racers are outstanding at what they do, and this marks the first time an autonomous system has beaten not one but two world-class human pilots.

To test the system, the UZH researchers set up a drone flight course (below). Both the autonomous drone and the human pilots were allowed to train on the course. Not only was the AI able to achieve the fastest lap time, but it also beat the two pro pilots through every waypoint by significant margins.

The AI uses external cameras to track the drone's path and make the proper calculations. The team hopes to modify the system to use the quad's onboard cameras. The use of onboard camera systems is vital for making other drone-related tasks practical. The researchers expect their work to be useful for applications such as search and rescue, building inspection, package delivery, and more.

The algorithm is also "computationally demanding." It currently takes up to an hour for the computer to precisely calculate the optimal trajectory. Because of this shortcoming, human pilots are in no fear of being replaced, at least for now. Clearly, in situations such as search and rescue, when time is critical, they will want a program that can more quickly calculate its path through waypoints.

All the technical details are outlined in the team's paper, which was recently published in Science Robotics.

Image credit: University of Zurich

Permalink to story.

 

kiwigraeme

Posts: 464   +360
We are told to be impressed - well I'm not that impressed - The World Champs have to get timings spot on - plus watching from a distance . All data is known except for minor tweaks - assume drones have build in sensors in.
Now if crosswinds, moving targets etc that would be more impressive

I mean I do not care about trains, planes and automobiles - but Jet Planes , Superbikes , Sports cars all have build in mechanisms to fly or go fast . Apparently super bikes really hard to ride real fast with out the CPU - again - I'm a distance 3 stages of separation observer.
A missile flies 3000Km - goes down the street into the open door -I know some are flown by remote pilots - but not all - The defense system kindly given to Israel takes out 100s of scuds flying fast
 

Fearghast

Posts: 417   +307
Even if it's the optimal, closed environment test, it's fairly impressive.
It will be interesting to watch, how will AI adapt to new variables when they will test it in some dynamic environment - wind, other drones, moving checkpoints etc.
 

Bobbydpue

Posts: 216   +145
Well I'm not that impressed
Apparently super bikes really hard to ride real fast with out the CPU
You likely haven't tried to build software and hardware to have a computer move itself through anything at any speed. Just like I used to tell my friends who were on the ground yelling up assistance to another climber "climbing is always easier from the ground".

Superbikes are easier to ride fast than other styles of motorcycles since that is what they were designed to do. While traction control and advanced ABS systems make riding superbikes safer they don't make someone a much better rider than they are. Riders still have pick the right points to start and finish a turn, still have to choose when and how hard to apply the throttle and how hard and when to apply the brakes. The onboard computers on any motorcycle don't make an impossible to ride motorcycle possible. It's just not how motorcycle riding works. What do you think the computers do on a superbike? Do you know anything about motorcycle riding? How did you form an opinion about something you know so little about?

Guided missiles like the ones you used as an example aren't flying through a circular course with obstacles to fly over, under or around. These missiles go from one point to another in mostly straight lines as fast as they are designed. These missiles don't fly near obsticles and they use satillite navigation to control their flight which isn't as sophisticated as controlling a high speed drone through a tight obstacle filled course.
 
Last edited:

Uncle Al

Posts: 8,155   +6,903
The greatest factor that stops the airlines from going to pilotless aircraft is the passenger. There is a calming effect to actually seeing, walking by and being greeted by the pilots when you board the aircraft. No doubt as the ridership goes younger there will be a gradual shift but for now I think the switch could be a disaster for the airline that tries it.
 

kiwigraeme

Posts: 464   +360
You likely haven't tried to build software and hardware to have a computer move itself through anything at any speed. Just like I used to tell my friends who were on the ground yelling up assistance to another climber "climbing is always easier from the ground".

Superbikes are easier to ride fast than other styles of motorcycles since that is what they were designed to do. While traction control and advanced ABS systems make riding superbikes safer they don't make someone a much better rider than they are. Riders still have pick the right points to start and finish a turn, still have to choose when and how hard to apply the throttle and how hard and when to apply the brakes. The onboard computers on any motorcycle don't make an impossible to ride motorcycle possible. It's just not how motorcycle riding works. What do you think the computers do on a superbike? Do you know anything about motorcycle riding? How did you form an opinion about something you know so little about?

I'm happy to educated - I claimed I knew stuff all about Moto GP - Sure I'd heard of Rossi who hasn't ? - but all those rider assistance have leveled the field to some extend - again plenty ignorant - assume speedway would ban it all - as it's the very essence of controlled sliding, drifting etc .
Ducatti or who ever control systems will feed real time data , and predict future tire traction etc . If I jumped on a superbike without launch control and open the throttle why stabbing the gears ( probably don't have std gears ) - I would most likely lose control. - all I saying is only the very best can ride on the edge without all the assists - I'm not detracting from their amazing skills and reaction times - correcting high speed wobbles etc - But all this processing is in real time , including predicting future lap speeds .
To the point at hand - AI drones probably have super accurate sensors ( accelerators etc )- and more control over power delivery - ie they cheat -( I assume as no robot controlling a RC transmitter . I mean that's what AI does - it starts bad and gets better with more runs . I assume the Moto GP , GP, all have very accurate simulators of the bikes and the tracks ( just guessing - however NZ defended the Americas Cup - and their simulators have to be a hell of lot more complicated than flying a drone on a set course with no wind, known air pressure , temperature , track layout , drones' ability ) ie in this drone racing - 1 hour before race all variables are known - and it's not some super very hard thing to predict like a 5 sun orbit system ). I mean Googles AI could learn Chess from just the rules in one day to beat the world champion - but not Texas Hold em - I'll still go with the Patriot Missile system as impressive
 

Puiu

Posts: 4,848   +3,736
TechSpot Elite
Well, we knew it was only a matter of time. I guess airlines will be switching over to piotless?
Airlines are already "pilotless" and have been for many years. The pilot is mainly there to watch the controls and system, everything else is done by the onboard computer (including landings). You can consider the pilot as the less capable backup system :)
 

Avro Arrow

Posts: 1,713   +2,022
TechSpot Elite
I have no issue with autonomous AI being used for civilian transport but I NEVER want to see them weaponised. A computer with really complex software is still just a bunch of algorithms and is essentially mindless. I don't like the idea of mindless killing machines nor do I like the idea of a psychopathic leader who imprints their psychopathy upon the AI itself. It's guaranteed that such a thing would happen because psychopaths are narcissistic and it would be a form of self-aggrandisement to them.

The first example in pop culture would be the Transformers. The Quintessons built an industry that manufactured autonomous robots that they sold like any other manufactured product. The Autobots were designed to be consumer-grade service automatons while the far more expensive Decepticons were the weaponised military models.

The Quintessons gave them advanced AI which eventually led to them first becoming sentient and then becoming sapient. As a result of their design and programming philosophies, the Autobots became a caring robotic race dedicated to serving the greater good. The Decepticons however, well, they had been programmed for war and conquest. It wasn't too long before all hell broke loose and the only reason that the Autobots weren't wiped out completely is that they outnumbered the Decepticons by about 10,000 to 1.

If you program an AI to be able to kill, that's exactly what it will do, if only to learn more about it. It's like having a child play video games like Duck Hunt and Hogan's Alley before handing them a .22LR semi-automatic pistol. Nothing good will come of it.

My personal biggest problem with weaponised autonomous AI is, how do you teach it the concept of surrender? Since computers are so task and objective-oriented, the concept of surrender, a very biological thing, would be completely incomprehensible to them. Even if you managed to teach it the idea of surrender, the first time it encounters a treacherous opponent who fakes a surrender to achieve a tactical advantage, it will most likely disregard surrender in the future, just like the good ol' ED-209 in Robocop:

WARNING - If you have never seen this clip (which means that you've been living under a rock), it's pretty gruesome:
 
Last edited:

Avro Arrow

Posts: 1,713   +2,022
TechSpot Elite
Well, we knew it was only a matter of time. I guess airlines will be switching over to piotless?
Believe it or not, these days, they more or less already are. The captain and first officer are just there in case something goes wrong. Commercial jetliners have been capable of taking off, following waypoints and landing (in various weather conditions) without human input for years.
It usually is a form of human/pilot error that causes crashes. Just like cars
That's a fact.
A missile flies 3000Km - goes down the street into the open door -I know some are flown by remote pilots - but not all - The defense system kindly given to Israel takes out 100s of scuds flying fast
This may come as a shock to you, but the Patriot system that Israel had during the First Gulf War didn't actually succeed in taking out Iraqi SCUDs. There was only one successful intercept out of 47 sitting-duck SCUDs. The media was told to engage in propaganda to the effect that the Patriot system was fantastic (when it didn't actually work) so that the US could sell it to more countries.

One missile that is actually effective at taking out other missiles is an advanced IR WVRAAM made by a European consortium led by Diehl Defence of Germany. It's known as the IRIS-T and it has a "missile shield" mode in which it can target and take down incoming AAMs. It can have trouble against WVRAAMs because of the small distances involved but it is extremely effective against BVRAAMs. It renders the idea of "first shot, first kill" essentially moot. The Raytheon AIM-9X also has this function although I have to assume that the IRIS-T had it first because I when I first heard about it, the AIM-9X was not yet known to have this capability.

One REALLY crazy thing about the IRIS-T is that some Norwegian software engineers fiddled with some of them and managed to get positive IR lock on ground targets which could make the IRIS-T the first all-purpose multirole guided missile (like the standard missiles in Ace Combat).

The ironic thing is that none of it really matters because none of this modern military tech is even remotely as scary as the atomic bombs from WWII, let alone the nuclear-tipped ICBMs and (especially) SLBMs that we have today.
 
Last edited:

mbrowne5061

Posts: 1,902   +1,105
So, while still technically impressive, I would not call this "beating" the professional pilots. The pilots had to operate their craft in real-time, while the drone had 1hr to calculate its path prior to flying it. Of course the drone is going to win. When it does it in real-time, then I would say the drone wins.
 
Believe it or not, these days, they more or less already are. The captain and first officer are just there in case something goes wrong. Commercial jetliners have been capable of taking off, following waypoints and landing (in various weather conditions) without human input for years.

Long time forum creeper first time poster. I felt compelled to respond to this.

I can't speak for all different types of aircraft but I have 9 years experience flying 737's and the aircraft it is not capable of taking off automatically. Although the aircraft can auto land, I can count on one hand the number of times I have done so outside of the simulator. Typically autoland is reserved for very low viability approaches. The autoland system relies on a ground based navigation system(ILS) and incorporates extra redundancies which are enabled during low viability operations to ensure the integrity of the signal. It is highly abnormal for it to be used unless necessary. Additionally, the system requires the pilots input to be properly programed and we are still actively configuring the aircraft for landing.

You usually have really good points but in this case I guarantee we are there for more than if something goes wrong.
 

Avro Arrow

Posts: 1,713   +2,022
TechSpot Elite
Long time forum creeper first time poster. I felt compelled to respond to this.

I can't speak for all different types of aircraft but I have 9 years experience flying 737's and the aircraft it is not capable of taking off automatically. Although the aircraft can auto land, I can count on one hand the number of times I have done so outside of the simulator. Typically autoland is reserved for very low viability approaches. The autoland system relies on a ground based navigation system(ILS) and incorporates extra redundancies which are enabled during low viability operations to ensure the integrity of the signal. It is highly abnormal for it to be used unless necessary. Additionally, the system requires the pilots input to be properly programed and we are still actively configuring the aircraft for landing.

You usually have really good points but in this case I guarantee we are there for more than if something goes wrong.
Well, I could definitely be mistaken as you're clearly more versed on the subject matter than I am and I appreciate that you usually find my points good. I'm not perfect and I do make mistakes sometimes. I do fly but I'm only GA, nowhere near an ATP like you and since landing is the most difficult part of flying (and it's still easy as pie), I had assumed that if a jetliner could auto-land that it could also auto-rotate on takeoff.

It doesn't make sense that Boeing, Airbus, Embraer, etc. wouldn't have that feature available as well but, if you're at 737 pilot, I completely accept what you say and defer to your experience. Thanks for the information.

I'd rather be wrong and learn the truth than be wrong and think that I'm right. I appreciate your input and envy the hell out of you for being an ATP who flies jets. I'm just a little guy who dreams of one day owning a Lake Buccaneer (to be honest though, a C172 is probably more realistic). :D
 

Markoni35

Posts: 1,228   +498
What do we say to "drone better than a pro pilot"?
"Not today".
The problem is that track was stationary. No computer vision needed. Just calculating pathways through predefined points.

Now, take a look at this video:

Especially part when it goes under the train. Not sure how could you even program something as dynamic as that, but first of all, the drone would have to see the train. Probably using lidar. Which would also enable it to compute distances and check if it can pass. Now, find me a drone that can do that using just the onboard camera. Just one camera, not even stereoscopic vision.

Then I can say, "it's better than a human pilot".