Champion-level Drone Racing using Deep Reinforcement Learning (Nature, 2023)
Вставка
- Опубліковано 20 сер 2024
- First-person view (FPV) drone racing is a televised sport in which professional competitors pilot high-speed aircraft through a three-dimensional circuit. Each pilot sees the environment from their drone’s
perspective via video streamed from an onboard camera. Reaching the level of professional pilots with an autonomous drone is challenging since the robot needs to fly at its physical limits while estimating its speed and location in the circuit exclusively from onboard sensors. Here we introduce Swift, an autonomous system that can race physical vehicles at the level of the human world champions. The system combines deep reinforcement learning in simulation with data collected in the physical world. Swift competed against three human champions, including the world champions of two international leagues, in real-world head-to-head races. Swift won multiple races against each of the human champions and demonstrated the fastest recorded race time. This work represents a milestone for mobile robotics and machine intelligence, which may inspire the deployment of hybrid learning-based solutions in other physical systems.
Reference:
Elia Kaufmann, Leonard Bauersfeld, Antonio Loquercio, Matthias Müller, Vladlen Koltun, Davide Scaramuzza
Champion-Level Drone Racing using Deep Reinforcement Learning.
Nature, August 30th, 2023
DOI: 10.1038/s41586-023-06419-4
PDF: www.nature.com...
To see more of our work on drone flight and machine learning, check out our webpage:
- Drone racing: rpg.ifi.uzh.ch...
- Deep learning: rpg.ifi.uzh.ch...
- Agile flight: rpg.ifi.uzh.ch...
- Lab's publications: rpg.ifi.uzh.ch...
Affiliation:
Elia Kaufmann, Leonard Bauersfeld, Antonio Loquercio, and Davide Scaramuzza are with the Robotics and Perception Group, University of Zurich, Switzerland: rpg.ifi.uzh.ch/
Matthias Müller and Vladlen Koltun are with Intel Labs.
Would love to see the full FPV footage. From start to finish. Also, would be great to show a side-by-side with the human FPV.
Maybe the human contestant could learn from the videos of the ai drone.
@@vladimirslavik5230 infinity reapetiting 😝 but at the end we f up
ua-cam.com/video/qEVx5t5l6Zw/v-deo.html
Just wow. I can't help but imagine a nightmare scenario where these little things hunt you down till you pass out.
An order of Swarming killer drones, coming right up!
Both amazing and terrifying at the same time! Makes me think about how no human has beat a computer in chess since 2005...
Quite. There is no question that the space where humans are still "the best" - or even necessary is shrinking every day.
This could be used for the good of all, but I'm expecting it will be used instead for more corporate profit.
or in wars :-( @@langdons2848 i mean, thats a subtopic of what you wrote: corporate profit. just with more grim outlook for humans.
@@langdons2848 even more advanced drones for war... 😬
@@langdons2848: That's pretty much guaranteed... Through the history of mankind, every time someone who intends to benefit us all comes up with something new, or an advancement of existing technology, there will be people who want to profit from it, and people who want to use it for evil!
Save the planet... don't invent things! 😜
@@nicholasfinch4087
So happy Google algorithm took me to this video. UZH's engineers are no joke.
splendid upload. 😍
Y'all didn't need to give the voice-over job to a computer too lmao
The voice over industry is on its way out. Sad to see it happen.
Dollars to donuts this YT channel uses Ai to compile every single aspect of video like this one
It’s not even a job, it’s just an unnecessary expense.
Every job that a computer can do,
A computer SHOULD do.
@@pr0xytv942not the kind of world I want to live in
i fly fpv race drones and i never even thought of the turns this AI is taking, amazing stuff
Blind split S do be hitting
So just like top chess players pick up tricks from AlphaZero, FPV racers can pick up tricks from this AI. Amazing!
this is very good news indeed. I hope #skydio colleagues will soon release retail version of their drone which will be able to follow fast-moving objects in tight environments like downhill bike track to make the best possible recordings in follow-rider-mode. even with several days of simulation and practice
A weapon of the future.
The Drone knows where it is at all times
great research about real-time DRL
They made a piece on The Independent about this today, starting with "a team from university of zurich in GERMANY"
a giant leap for the robot, and the human
awesome!!! It is Nature!
I welcome our new AI drone based overlords! Please let me live in peace in my cabin in the woods. Thank you. :D
I wonder how close the AI was to always flying the exact same path?
Maybe that's why they have movable gates.
The inherent uncertainty of the environment makes this very unlikely
You mean reapetiting the perfect path
Great Job nice !
so cool
Awesome invention in RL and Vision
Amazing video and research
Wow! That's swift! 😯
Auto pilot flying cars are going to be lit
This is insanely good news ❤ champion work guys 💪
OOOOOOOOOOOOOOOOOOooooooooooooooooooooooooooooooo
Amazing.
Awesome 😎
thank you from manhattan ©2024
Awsome
Please for love of god bring back ADR (Autonomous Drone Racing) back to IROS in 2024. Last one happened atleast 4 years ago, it will be very interesting and competitive now
train a human on a simulated approximation of the course for 1 hour and i'm fairly sure the outcomes would be similar. still, amazing abilities in removing latencies to be capable of physical reaction so rapidly and excellent process optimization, for drone users in a rush
Wow
Skynet will be pleased.
Next step: Slaughterbots. What a time to be alive!
On day we wont need an hour of compute to freestyle!
this shit about to blow up
currently at 99k
Wasted opportunity to name it "Scar" - as a homage to Battlestar Galactica. Because that's what they are making :P
How would a change in air condition affect this? eg change altitude, or pressure, or humidity? I imagine each to affect flight dynamics. Would the system know/adapt?
Good question. I noticed that they didn’t show side-by-side racing. Can it handle turbulence?
amazing!! so what would be the next to-be-conquered limitation? would it be the ability to process the images of different environments?
M 7
I assume it would be removing the pretraining step for each track. Instead you'd fly one lap manually to teach it the gate order and then let it loose.
Satisfy your wife.
@@Bruno-cb5gk i wouldn't go this far, since human competitors also don't start from scratch. What I would like to see is dynamic adjustment of the residual net to the current environment (active learning).
now it just has to get consistent :D
Only time till its found on military operations
How does anyone know which gate to go through next? That is the most confusing race track I've ever seen.
Just like how any other racer learns a track. Study and practice.
You think rally racers jump onto a track blind? 😂
@@pr0xytv942 I prefer tracks that are clear and new. Something none of the competitors have seen before the race. Then the competition is based more on skill instead of memorization and repetition.
Do NOT let the Chinese get hold of these algorithms...
What is the generelization capability of the system? If it goes off the learned path distribution, can it recover into it's learned path?
Can it beat MCK both with 1 hour training on a fresh course? Highly doubt it. Still cool though 👍
You also have to ask the question; Can someone off the street who has never flew a drone before, reach MCK level within the same time frame as the AI did? I very highly doubt that either.
A MCK need years of training in order become that good.
You're not very good at extrapolating into the future are you.
It had a 60% winrate. What are the conditions that caused it to fail the other 40% of the time? From the article I read it could not handle being bumped at all. What else? Also, training on a very specific track for 23 real-time days of running to calculate an "ideal pathway" through a virtual, fixed course really makes this not much different than automating a speedrun of a videogame by preprogramming inputs. Does a human champion practice this exact course in an accurately recreated environment for 23 days? I highly doubt it. Seems like an exercise in "specific" vs "general" and the obvious interest in "being the first to beat humans in a real-time sport" is absolutely disingenuous.
It's better to ask, is a typical human with really basic flying experience able to somehow reach champions' level in 23 days practice? Right now we see how technology is raising. It's not so far to a point that AI only needs to scan a track, does some calculations and it's ready to fly faster than MCK, who needs to at least fly a track a few times before challenge.
The impressive thing is that it's able to fly with speed and precision despite being fed very noisy data.
They should have their own separate league.
The competition is who can program their drone to auto pilot the course to the win.
Humans are manually flying when competing with each other.
@@wKichwell. The simulated time will be more than 23 days by far.
@@rubencid2575 do we count all fpv related experience and time to learn/practice on humans? because ai learn from scratch, they don't have any background on start
is the code somewhere available - i would like to study it
aLeX vAnOvEr wOrLd ChAmPiOn, bring MCK next time.
Maybe you should try with random moving targets, that would be interesting. But at this video its similar to play chess against a chess bot, it has near-optimal solution and basically you are trying to keep up that solution as a human.
Humans wouldn't fair well either. They train on fixed targets as well
Was the algorithm / RL was trained on the exact course the drone completed?
yes, but the approach is general. So if you were to employ this trained model on another track it wouldnt work right, unless u retrain the model from scratch, using the same approach.
Should have had mck fpv there hes the real robot 😂
I wonder if you change circuit dynamically, what will be change in results?
It would have to be re-simulated for the new course setup
@@TheRedstoneFactory who will be the winner?
So, can this company give me drone software to help film me while I mountain climb/scramble?
very fun, until it kills...
Cyberdyne systems Skynet
dang china.. why isnt this in the us :(
let me guess most of the comments:
"terrifying"
😑
useful for the next war (as drones are already in use to destroy humans) 😞
my first thought as well
4:34 what about chess and Go or video games? are they not real world competitive sports?
"real world" ig refers to wherein you stuff moves physically and, i mean not board games & video games
Not Fair.
Calculations must be based only in the same imputs as the human... the video on the VR.
AI has an IMU that the human doesn't have.
Scary to think it will be impossible to escape the future killer drones powered by AI.
I wonder whether the AI is embedded in the drone aka autonomous or network based. A network based AI would surely suffers higher latency. What is the exact latency of a network based AI both in terms of transmission and computation? What if the bitrate dropped and there is a degradation in the network? This is a common issue with FPV drone especially when dealing with long range and signal penetration through dense environment.
They make clear : _"on board computer, single camera and an inertial sensor"_ (0:50) ua-cam.com/video/fBiataDpGIo/v-deo.htmlsi=nUlR4RPi_W2mHVis&t=50
They specifically said the training is done on different hardware, but the actual piloting AI is entirely onboard.
So the drone is heavier due to the gpus or wtv required? How does it still perform better?@@fxm5715
really, not very surprising, is it? put differently: what took you so long ;-)
@Ukraine
pool playing robot did it first, no?
Go and kick one of the obstacles a couple meters to the side. See what the "AI" does then....
That was my initial idea as well
Trains again (like any other pilot would do to achieve the best times) and fly it afterwards like shown?
Yeah, sounded like - at least based on the video - that they trained specifically on this track. Would be more interesting with a more general model.
@@Convergence33 do you think any pilot first will train? Or does the pilot use his brains? If you move stuff around, change shapes, pilots can do. This needs training like you said. It can only train from the past.
The article I read accompanying this states that the drone has massive limitations that are effectively not feasible for "winning" in real world conditions. For example it was trained for 23 days virtually. Is that how these courses actually work?
NGL it's probably time to collectively stop improving AI so we don't go extinct.
There's open source AI. It's like trying to recover dust in the wind. There's no stopping it now
Sell it to the military if they don't already have the tech😂
Snake oil! Take your system out in the real world on a track that it never seen before training and spar that against a human. I can tell you who will win that challenge.
So the drone is not real time? Its pre programmed to fly in a specific way? Thats cheating.
as the "autonomous" drone calculates the whole track in advance ... its cheating.
Unwatchable - using a computer generated voice which is intensely irritating.
Clearly you guys put alot of work into this. Couldn't you take 4 more minutes to actually read the script yourself? People are sick of these ai voices
1. start. - AI's response to the start command is surely faster. this is what your "-0.5 second record" is. provide individual flight of each participant with the TIMER to START at the moment of TAKE-OFF, not at the moment of giving a start command.
2. it is a trained track. provide comparison to humans on new, untrained tracks.
ONLY THEN try to bulshit me that AI is better again.
One week training for the humans is how many attempts? A few dozen? One week training for deep RL is... thousands? millions? of "attempts". Not even a close comparison. How about they give the deep RL only 24 iterations of its learning algorithm, and we'd see who'd win. These kinds of comparisons are totally misleading and just clickbait. That it got into Nature is incredible[-y dumb].
they should race against @MrSteeleFPV ;-)
Steele is a freestyle pilot not a racing pilot. He will absolutely get rekt by the 3 racing league pilots easily
All three of the pilots mentioned are faster than Steele. He is a freestyler who occasionally makes cinema content. An amazing pilot, sure, and better than I could ever hope to be, but he is not a racer, and doesn't pretend to be.
@@Rixynator what's wrong with a freestyle course...would be interesting what AI can do there, too. At the end of the day, humanity lost....sad but probably true.
So, when you move, remove or add the obstacles, the drone won’t able to reach this scores? It needs to retrain in mew environment? Does this even make sense, lol?