Autonomous drone defeats human champions in racing first
283 comments
·June 4, 2025bri3d
generalizations
From near the bottom:
> One of the core new elements of the drone’s AI is the use of a deep neural network that doesn’t send control commands to a traditional human controller, but directly to the motors.
bri3d
I saw that too - I'm assuming it means they're indeed using the DNN for stabilization. This has been done several times over the years, but generally with results which only rival PID and don't surpass it, so that's quite interesting. What's odd is that the physical architecture of the drone doesn't really make sense for this, so there must be some tweaks beyond the "spec" model. Hopefully some papers come soon instead of press releases.
sorenjan
They reference ESA's research in "Guidance and Control Nets", and when looking at ESA's page for their "Advanced Concepts Team" [0] they in turn reference ETH Zürich's research in RL for drone control. Specifically [1] this paper from 2023: "Champion-level drone racing using deep reinforcement learning" [2]. They use a 2x128 MLP for the control policy.
[0] https://www.esa.int/gsp/ACT/
[1] https://www.esa.int/gsp/ACT/projects/rl_vs_imitation_learnin...
koolala
This is crazy, its dexterity and range of motion could potentially exceed all human modeled systems.
HenryBemis
I assume that they shave off milliseconds by doing so, and a gyroscope (or similar) sends back the position/angle of the drone. And like this does it bypass the 'limited' onboard computer and instead uses a much better/faster computer?
pjc50
Reports downthread suggest that the NN is running directly on the drone, in the form of a Jetson. Which would give much better latency and quality of video.
itishappy
There's a few more details in the press release from the league itself. Sounds like they were really trying to put these things through their paces.
> The course design pushed the boundaries of perception-based autonomy—featuring wide gate spacing, irregular lighting, and minimal visual markers. The use of rolling shutter cameras further heightened the difficulty, testing each team’s ability to deliver fast, stable performance under demanding conditions
https://a2rl.io/press-release/9/artificial-intelligence-triu...
NegativeLatency
I imagine the slower speed is a closer fit to combat drones (which have a payload and sometimes a fiber optic cable)? Also watching MultiGP they sorta move/accelerate too fast for me to fully appreciate the maneuvering.
Feels kinda similar to the innovation around manned aircraft about 100 years ago when we went from toy/observation platform to killing machine in only a couple of decades. With the ardupilot news today, it was hard to not watch this and imagine the applications to a combat environment.
close04
> which have a payload and sometimes a fiber optic cable
The optic cable is for the human pilot. An AI piloted drone doesn't need it.
akie
If we are ok with AI drones autonomously choosing bombing targets, then you're right.
closewith
Although even autonomous combat/ISTAR drones may require fibre spools for BDA, ISTAR, etc.
Aurornis
> I imagine the slower speed is a closer fit to combat drones
A lot of comments are trying to draw connections to combat drones, but drone racing like this has been a hobby thing for a long time. The capabilities of the drones are set to have an even playing field, not to match combat drones or anything.
These aren't meant to have any parallels to combat drones, drones that fly long distances, or drones that carry payloads.
It's really just a special-purpose hobby thing for flying through a series of gates very quickly. Flight time measured in a couple minutes, no provisions for carrying weight.
david-gpu
We all understand that. People are simply observing that there an obvious path from this technology demonstrator to something similar in the battlefield.
pacetest
It used a small RL trained network running on the flight controllers MCU directly that controlled the motors given state (position, orientation ...) inputs. The Jetson handled vision processing.
wepple
> The spec DCL drone isn't very capable compared to the more open-specification drones in racing leagues like MultiGP
Yeah, I’m sure this is a great milestone but it isn’t notable until AI is beating MCK[1] who would be the “Lee Sodol of FPV”
durandal1
Also, this track looks nothing like a competitive drone race track, the obstacles are easier and it seems designed to cater to the autonomous drones.
bri3d
It’s DCL, they’re all kind of poor like this IMO. Certainly nothing like MultiGP
null
walrus01
> which is present on the drone and use direct IMU input and ESC commands from the Jetson
Does the jetson board even have the appropriate UARTs on it to talk directly to the ESCs? Your typical hobby grade 5" class size ESC (either 4-in-1 or discrete) cannot talk to two different controllers at the same time. If it's already wired to the four UART outputs from a 30x30 size flight controller (such as something STM32H7 based running betaflight), the ESC cannot be in communication with any other device.
jandrese
This is only a few days after the massive drone attack in Russia. Only a matter of time until we have drones smart enough to dodge bullets (or at least dodge out of where guns are pointing) while flying at breakneck speeds being controlled by AIs we don't fully understand.
The tech industry is working hard to bring about the Terminator future.
allturtles
“What hope can there be for mankind,” I thought, “when there are such men as Felix Hoenikker to give such playthings as ice-nine to such short-sighted children as almost all men and women are?”
And I remembered The Fourteenth Book of Bokonon, which I had read in its entirety the night before. The Fourteenth Book is entitled, “What Can a Thoughtful Man Hope for Mankind on Earth, Given the Experience of the Past Million Years?”
It doesn’t take long to read The Fourteenth Book. It consists of one word and a period.
This is it:
“Nothing.”
--Kurt Vonnegut, Cat's Cradle
fsloth
Vonnegut’s bleakness is not theoretical which gives it a specific bite. As POV Vonnegut cleaned up the shriveled remains of civilian victims of firebombing.
randomtoast
And Great Britain just announced plans to deliver 100,000 of them to Ukraine. Ukraine lacks the manpower compared to Russia. It seems logical to strengthen their forces by deploying these flying mini terminators. I believe we are not far from large-scale drone warfare. In World War II, we had epic tank and aircraft battles; now, the time has come for autonomous drone battlefields.
pjc50
Missed that news; https://www.reuters.com/business/aerospace-defense/britain-p...
I think we're already deep into large-scale drone warfare. Destroying a third of the enemy heavy bomber fleet is pretty substantial. It feels to me like that attack operated like Pearl Harbor, a marker that the old way of surface naval warfare / air attack was being replaced by a new one.
Don't forget that Russia has their own drones. They were the first to deploy the fiber-optic cable drones as an anti-ECM measure. And of course both sides are ordering parts from China.
IshKebab
$4k per drone... ouch.
gusfoo
The UK has also released some details of a new 'in bulk' RF-based drone takedown DEW. https://www.youtube.com/watch?v=F1vY5efYXMQ
ponector
100k kamikaze drones are not that many for the current stage of war. Millions are consumed annually.
jajko
[flagged]
NoOn3
I'm afraid they won't have a choice. Without Russia The West and Nato will crush China fast.
MetaWhirledPeas
I hope this scares nations away from warfare altogether, given the unpredictability of outcomes with heavy drone usage. Even China: they may have a lot of the little buggers, but that Ukraine attack took a lot of scheming. I doubt China can out-manufacture human and AI trickery.
bluealienpie
That future arrived years ago. Now it's targeting journalists, women and children. Ukraine is just doing it at a lower cost.
greenavocado
Let's review what Uncle Ted had to say about this.
See paragraph 87 by searching for "THE MOTIVES OF SCIENTISTS"
https://dn790000.ca.archive.org/0/items/IndustrialSocietyAnd...
cess11
The conclusion is succinct and the stuff leading up to it of dubious quality.
"92. Thus science marches on blindly, without regard to the real welfare of the human race or to any other standard, obedient only to the psychological needs of the scientists and of the government officials and corporation executives who provide the funds for research."
nothrabannosir
Obligatory link to the short film (future documentary) “Slaughterbots” (2017), which depicts exactly this in harrowing detail:
insane_dreamer
This is portrayed in Ministry for the Future which describes AI controlled swarms of small drones/bombs that fly apart and come together at their target and are almost impossible to stop.
chrisweekly
Fantastic book, highly recommended.
Taek
I tried to read the book and to me it came off as little more than doomer and disaster pornography. I found a lot of the situations to be far fetched and didn't feel like it portrayed a realistic image of how the world works.
null
trhway
>drones smart enough to dodge bullets
well, there will be similarly smart "predator"/defense drones. The humans will have no chances on such a battlefield populated by thousands drones per square kilometer fighting each other.
>The tech industry is working hard to bring about the Terminator future.
And i think removing people from the battlefield is a good thing.
>or at least dodge out of where guns are pointing
just a bit of arithmetic comparing new weapons - drones vs. classic guns. Say a radar guided gun takes 1 sec. to train onto a drone and shoot several bullets. The range is max 3 km (an expensive 20mm-30mm autocannon like Pantsir) - 35 seconds for a 200 miles/hour drone. Thus all it takes is maximum 36 such drones coming simultaneously from all the directions to take out that gun. At less than $1000/drone it is many times cheaper than that radar guided gun. (and that without accounting for the drones coming in very low and hiding behind trees, hills, etc and without the first drones interfering with the radar say by dropping a foil chaff clouds, etc.) It is basically a very typical paradigm shift from vertical scaling to horizontal scaling by way of software orchestrated cheap components.
dragonwriter
> And i think removing people from the battlefield is a good thing.
Drones don't remove people from the battlefield, they further the trend of there being no boundary to "the battlefield", putting everyone on it.
They can, depending on how they are employed, reduce the casualties (total and particularly civilian) on both sides of a conflict for any degree of military impact (Ukraine's recent strike against Russian bombers is an example), or they can increase the civilian death toll for marginal military impact (the accounts of Israeli gun- and missile-armed drones directly targeting civilians in Gaza being an example of what that could look like.)
impossiblefork
>And i think removing people from the battlefield is a good thing.
It is very dangerous, since it will mean that an organization with enough drones can dominate society on its own. Much better if humans were battlefield-relevant.
trhway
It is understandable pure-logic thinking until you're the one to be made battlefield-relevant.
And if you look at Russia your logic does fail on that example - no amount of human losses affect Russia's behavior in the current war as they are sure that Ukraine will run out of soldiers before Russia does. So, from Russia's POV the faster the grinder the sooner their victory.
MangoToupe
Don't we already have this in the form of the state?
throwaway2037
About your last paragraph: High level, I generally agree, but when you dive deeper to look at the numbers, I doubt that you can have 1000 USD drones that can fly at 200 mph/320 km/h. One quarter of that speed, I could believe.
trhway
quad - no. Plane type - very doable.
cess11
"And i think removing people from the battlefield is a good thing."
You're mistaking the removal of certain soldiers for "removing people". There will absolutely be people in future battle fields, mainly civilians, or as we call them now, terrorists.
paganel
> And i think removing people from the battlefield is a good thing.
I agree with your other points, but this only helps with (physically) extending the battlefield, at least going by the current war in Ukraine. It's not only the line of contact that is now part of the battlefield, there's also a band of 10-15 kilometres (if not more) on each side which is now part of the active battlefield because of the use of drones.
Even though I have to admit that it looks like the very big power asymmetry in favour of cheap drones over almost everything that moves down bellow (from mere soldiers on foot to armoured vehicles) has helped with actually decreasing the number of total casualties (just one of the many paradoxes of war), as it is now way too risky to get out in the open so soldiers do it way less compared with the pre-drone era.
gattr
I agree, but I'm a bit disappointed it will probably come to this, instead of having a mano a mano like in the movie "Robot Jox".
pjc50
+1 for mentioning that movie; I watched it a month ago and it's hilarious. Nearest I've seen to live action with giant robot anime sensibilities.
GenshoTikamura
[dead]
atonse
Oh man, can anyone imagine a non-Terminator scenario for this?
Update: I'm not saying people shouldn't develop this, we're never going to squash human curiosity. But when I see this kind of stuff, I'm deeply troubled by how bad actors (state and non-state) will use this.
I hope our security services are working hard on countering these potential threats.
AngryData
Im more worried about these type of things causing us to blast each other and ourselves back to the 1920s or so during conflicts when small explosive EMPs start being viewed as less damaging than drones and robots. A fast explosive on the back of a neodynium magnet and a few coils of copper can make a hell of an EMP blast. The only reason we don't use them now is due to all the collateral damage, but if drone bombs represent even more damage they become viable. Yeah it will destroy all the radios around and fuck up a bunch of expensive equipment, but you would still have soldiers with guns rather than just smoking craters.
pjc50
> A fast explosive on the back of a neodynium magnet and a few coils of copper can make a hell of an EMP blast.
I'm having a hard time believing this is effective.
> The only reason we don't use them now is due to all the collateral damage
Russians don't care about collateral damage and there doesn't seem to be any evidence of them using such weapons?
AngryData
Nobody really uses much undirected EM warfare in my opinion because it represents a huge escalation in a war, similar to the use of indiscriminate chemical weapons, or even nuclear weapons.
It would be devastating in the local battlefield, potentially frying radio or other equipment depending on the size of the device or how close you could lob it towards the enemy before going off; but with the low wattages many non-military communication devices use today you would also be blasting horrible noise to all of them beyond the local area and disrupting communications across potentially multiple neutral countries.
It would be a large act of aggression against any countries around them and NATO, and at scale possibly even piss off far away countries like the US and China. Especially large EMP devices could even be temporarily misidentified as a nuclear explosion and gain the immediate full attention by any nuclear powers watching out for it.
Legend2440
You could do EMP, but you could also do some sort of point-defense turret. Drones are lightweight and fragile, so it doesn't need to be big - just fast and auto-targeting.
marcus_holmes
Didn't they try this in Ukraine and it doesn't work? Any point installation is quickly overwhelmed. The only answer to FPV drones so far seems to be more FPV drones. Though they're not using fully-autonomous drones in Ukraine yet, so that might still play out.
jmccarthy
very prompt burrito delivery?
cluckindan
If by burrito you mean shaped charge high explosives with lethal shrapnel, triggered by facial recognition, delivered by drones the size of house sparrows at the speed of sound, then yes, burrito delivery.
roughly
Christ, you sound like my nutritionist.
generalizations
In china probably very soon. In the US? Regulation has already killed that.
lbotos
I feel like search and rescue after an earthquake where a drone swarm can canvas and categorize if it saw movement or not is one possible "non-bad" use.
martin_a
Fire departments and police in Germany are deploying more and more drone units, too.
Firefighters use them to search for missing persons but also to get aerial images and a better overview of larger scenes as "running around" is often not possible or doesn't help that much with the overview.
Police is using them to take pictures of accidents. It's easier to see tire marks and the whole "history" of an accident from above. Really reduces their time on a scenery to take pictures of everything.
energy123
Paul Christiano has thought about these scenarios. I recommend his interview with Dwarkesh a while ago where he goes in depth about it.
MoonGhost
This will definitely be used in drone vs drone dogfight. Interceptors hunting spy, bombers, and kamikaze drones.
contravariant
Sure, just strap a nuke to it and watch WWIII kick off. No terminator necessary.
AlienRobot
Drones flying through your windows to deliver things faster.
Cons: massive invasion of privacy and probably illegal.
Pros: looks cool.
itishappy
I've always thought a user-installable drone-pad in the style of a window AC unit would be the ideal.
TYPE_FASTER
Inspecting utilities and other industrial infrastructure.
DoingIsLearning
The actual race is also worth watching:
https://www.youtube.com/watch?v=yz2in2eFATE
The speed and flawlessness is quite impressive considering it is being resolved with what I imagine is noisy inertial data and a motion blurred CCD camera.
zellyn
Looks like most of the comments here are about the use as weapons and the possible dangers. I believe "Slaughterbots" is the canonical sci-fi video on the subject, and it appears to be aging pretty well. Unfortunately…
sveinatle
I remember being blown away by a TED talk were "minimum snap trajectories" are planned for quadcopters to fly through hoops and slots.
It's really cool to see this happening fully autonomously and at such high speed. I wonder if the use of AI means that the approach is fundamentally different, or if it uses the same principle of minimizing snap?
https://www.ted.com/talks/vijay_kumar_robots_that_fly_and_co...
pacetest
It's fundamentally different, it's using an RL trained network that gets the drone state (position, orientation, velocity) as input and directly outputs motor commands.
airstrike
Interestingly, the URL for the embedded youtube video ends with the word "FATE"...
rossant
Gives me the idea for a silly game: finding YouTube videos with words in their identifiers that are relevant to the content.
_ache_
Nearly two years from comparable to human to beat the best.
https://www.npr.org/2023/08/30/1196777528/an-ai-quadcopter-h...
ilikeatari
Looks like it had NVIDIA Jetson Orin NX 16 GB. No GPS, Lidar, motion capture so its vision only. 6s battery so 5 incher?
ilikeatari
Does anyone know the FC or AIO they are flying?
Leo-thorne
I’ve seen AI beat humans in simulations before, but doing it on a real track with the same hardware is honestly kind of amazing. What surprised me the most is they didn’t use any traditional flight controller. They just let the neural network handle the flying.
I’m really curious how this would perform in messier, less controlled environments.
shinycode
A side note, will we still attend and watch Formula 1 races it AI would drive cars (maybe near perfection) ?
tialaramex
Human sports remain interesting when the humans are notably worse at whatever they do than a machine purpose built for it, or indeed wildlife that specialised for this.
Usain Bolt was the fastest human sprinter in the world, but compared to a good motorcycle over paved road he's obviously not very fast, and likewise compared to an emu. Nevertheless, Bolt's 100m performance drew big crowds, even though people also watch Motorcycle racing and (I think?) Emu racing.
It's like speed running, the categories are arbitrary and self-selecting. Why the Modern Pentathlon? Why not. Why Super Mario Warpless? Why not. If everybody wanted to do Super Mario, only the odd numbered levels and also you must kill all the enemies, that's what the run is, our choices are arbitrary and we value whatever we like.
frakt0x90
Yes. People thought computers would kill chess, but despite current chess engines being able to trounce every human in existence, chess is more popular than ever.
zemvpferreira
That said “regular” chess is deeply in crisis, with less computer-assisted formats coming up to challenge it.
mrheosuper
similar to drag racing i guess.
miki123211
This makes me wonder what the "best" vehicle for human racing would be like, if there was no requirement for a human driver.
Let's say your task is to move a human from A to B (by a pre-designated route) as fast as possible. The only conditions are vehicle weight, no outside radio contact and no damaging the road (assume each vehicle goes separately, so e.g. slip stream effects don't matter). You can rely on the human to drive or use AI, you can go on the ground or fly through the air, anything is allowed. What would be the best way to do this?
83
>> The only conditions are vehicle weight, no outside radio contact and no damaging the road
Desiccate the human and compact him into an aerodynamic shape. Carry the (now much lighter and more aerodynamic) human inside a small rocket.
polonbike
Well .. it depends on other assumptions, like the amount and type of energy allowed (continuous gigawatt electricity access during the trip ?), amount of "roughness" the human can sustain during the trip (canonball a human packed in an big airbag ?), actual budget limit for the project, etc ....
npilk
Not exactly what you're after, but similar: https://what-if.xkcd.com/116/
1970-01-01
It looks like the reason it won is due to a human pilot crashing into the obstacle. Not exactly an unfair problem, but worth mentioning that it was close until human mistakes were made.
This is quite cool since past efforts in this direction have usually relied on crutches like outside-in imaging and positioning.
A few details I picked up:
* The drones are a spec drone across the league. It's a fairly large-footprint FPV racing drone (it's a 5" propped drone, but it's very stretched out and quite heavy) with both a Betaflight flight controller and a Jetson Orin NX onboard. Teams were only allowed an IMU and a single forward camera.
* It's unclear to me whether the teams were allowed to bypass the typical Betaflight flight controller which is present on the drone and use direct IMU input and ESC commands from the Jetson, or whether they were sending and receiving commands from the flight controller and relying on its onboard rate stabilization PID loop.
DCL is kind of a weird drone racing league since it's made for TV; it's mostly simulator based with, more recently, only few real events a year. The spec DCL drone isn't very capable compared to the more open-specification drones in racing leagues like MultiGP, in large part to keep the events more spectator friendly. This probably makes it more amenable to AI, which is an interesting side effect.