Hi everyone! This project is currently very limited by funding as I'm a broke high school student lol. If you'd like to support it, please consider donating at buymeacoffee.com/asrocketry .
@ASRocketryamazing work. I work in the aerospace industry. I suggest looking into finding a top tier university and contacting the directors of programs you’re interested in. They might be able to help you with your passion projects and in the future scholarships. Amazing work. Keep it up!
This is absolutely incredible. The fact that the model is trainable on local hardware, in a phone app and then directly uploaded to the fcc is a genuine paradigm shift in consumer/hobbyist control systems. What you have here is truly something special and you should be immensely proud of yourself. Much love from Northumberland.
Using LSTM or just temporal data is a really nice way of not only getting rid of the need to train your NN whenever the dynamics of your rocket changes, but also for closing the sim-to-real gap. We use this with our autonomous quadrotors, where the control system can adapt to different dynamics based on a few seconds of flight data. This also helps a lot with external disturbances like wind, that aren't being simulated. You can also run 2 networks, one that estimates coefficients representing the external disturbances (forces & torques) based on temporal data, and then the second network that does the control can leverage these coeffs.
sounds interesting, the ascent of these flights only actually lasts about 2 seconds though so there's not really much time to make corrections like that. My simulation currently has many types of disturbances simulated, but I'll have to look into that for detecting things like slight motor misalignments which are currently a pain
@ASRocketry I'd be inclined to consider a policy optimizer solution like PPO to pretrain a model in a simulator..... and then use real world data over many flights to post-train/fine tune. TVC isn't an inherently difficult problem to optimize for, thus the reward function can be really quite straight-forward and will probably converge quickly.
Im a Computer science student, and fascinated by the AI aspect. A deeper explanation on how you set up the testing environment for training would be very cool :)
Very cool. You used a genetic algorithm to evolve feed-forward neural network weights using basically PID inputs, this allows non-linear correction. If you REALLY want to overengineer this, one thing that imo would be really cool is to essentially set up a 1-d convolution or even a LSTM on the input - then, add some noise to your angular velocity and angle measurements at random, add some dropout in the weights and you'll get something that is robust from somewhat noisy sensor data. Update your training to include a bunch of different possibilities during launch - burst of thrust, very high wind gusts, different angles of initial launch conditions, etc and it becomes even more resilient. Keep up the awesome work!!
Lots of types of noise have been added to the simulation which allows it to cope with strong winds already! The day of the launch shown in the video people’s chairs were blowing around lol
@@AndrewOrtman Thank you! And thanks for the huge donation thats insane lol. I currently dont have a discord but I might make one at some point - I think there's an option to post to supporters on buymeacoffee so I'll probably put some progress updates there between videos
This is incredibly impressive. I’d love to see a video going into more details of the various aspects, especially a step by step guide to how you did the AI stuff. Liked and subscribed!
Thank you! I'll think about doing it but it will be a lot of effort lol, however I am enjoying using manim a lot. I've found a way of making the network a lot more powerful so I might merge both things into one video.
I was going to type the same thing about more details. It’s very interesting and your channel could really help other hardware engineers use AI for their projects if you do more step by step videos. I for sure I would love it. Great effort👌🏾
Well done! I look forward to the evolution of your ideas. Your story is well-structured, and your approach based in first principles - a great combination for RUclips.
This is awesome! I’ve dabbled in TVC stuff before but always used a traditional control method. I’ve always wondered about using some sort of machine learning, and you went and nailed it! This inspired me to go give it a whirl as well! Great work
This is insanely cool, you train the AI sort of like they do in trackmaina. Not surprised it worked this well, many concepts are surprisingly similar. What do you use for the engine? Just a mounted firework? And simple servos to point the engine? This rocket seems like a very affordable yet cool project.
Thanks! also a big fan of trackmania lol. In the UK rocket motors are annoyingly well regulated so these are specially made for this. I wish it was cheap but hey I got there eventually :)
I would love to learn more about the ai training/simulating process. The on-site training could also allow you to embed accurate wind speed/ gusts into the simulation
This project is really impressive! By the way, I'm curious about the details of the rocket simulation. I hope you'll make an explanatory video about it in the future!
The quality is awesome! Excellent job! I don't know anything about automatic control systems, so I don't understand the difficulty of tuning their parameters. I bet that if we had perfect sensors and instant responses to changes, the robot could perfectly interact with the environment. But we probably don't have either. So, what is the most difficult part of such a system? Is it the metrics fetched from the sensors, the delay of the response, the randomness introduced by the environment, or something else?
BPS described the challenge once very well as it being like "balancing a broomstick on your hand while blindfolded and wearing an oven glove" lol. It's basically just trying to keep a rocket upright when you've only got a single point of contact from the thrust of the motor
That worked beautifully, I wish rocket motors could run 10 times longer without aditional size/weight. Would love to see how it performs over a longer time. Great job man 👍
I'm fascinated by your project. It's really cool! Can I learn more about the on-board computer? I'm doing this myself now, and I'd like to look at the files.
@@YetAnotherPilot I’m not sure how you’d compute the losses for back propagation, this runs in only a couple of seconds anyway lol. It’s much easier to evaluate the score of a whole simulation than anything else and then there’s not really a need for anything other than random mutations until it works
Hey, I developed a joint that might work great for rockets...I always wanted train a neural network on that joint but never got the hang of it! So more info on this would be very helpfull! Thanks for the great content! 😊
@ASRocketry I build an SPM joint, all the files are available if you'd like to build or improve it! I am currently using metal bearings which are making that joint pretty heavy, but I am planning to build a lighter version soon!
Oh my goodness! Let me say that this is both an incredible project and a very well made video. You have so much talent to be able to do this prior to studying at uni. You have such a bright future ahead of you.
@ASRocketry yeah thats what I keep running into, the PID or scheduled PID's are just so easy and everything else is so hard. But what you've done is amazing.
@@testpilotmafia862 my professor of signal processing knew prof. Kalman personally. But when he shown us whole page of equations, he explained that by "you know, it's simple, just ... you know.. you put some number there and there... it's simple... just try..." Nobody ever understand that.
Absolutely wonderful creation, I am in tears! To further improve performance, start using more sophisticated neural networks that are designed for real-time/time-series operations. You can start out with RNNs and CNNs and their more advanced descendants like LSTMs, attention mechanisms, and Resnets :3
I love the implementation of the real Neural Network and not the "AI" scam (LLMs..., booring) - keep going, I did remote turret once but without NN, it's on channel, maybe it's time to revisit the idea 😃 To the Moon!
It'd be interesting to see the pid graph in comparison to the network. Specifically I wonder how it would do if you modified the pid settings using a network, vs doing this. Since the time aspect is built into the angular velocity, it'd be interesting because with the layer setup you showed you should be able to essentially fully represent the PID equation in it. So it'd be cool to see if that's the case by both seeing how many neurons are barely used in the hidden layers / how the performance compares
I think it does effectively do something similar to pid, and yes I could've done auto tuned pid. But... there's stuff coming that uses neural networks in another way so I was just using this as a starting point 😉
I honestly just free-styled a lot of it… which is probably why I ended up doing things like fixed topology. I used to watch carykh evolution videos a lot so that’s probably the source of most of my knowledge
I wanna know the basics of making a rocket with Raspberry pi chip and tools that I need to buy to make a rocket in my home. Please make a guide video on it. I would be very thankful to you
cool project! Question: what do/did you study? What do you say is necessary to learn to get into that stuff? I've studied electrical engineering but I don't feel comfortable with mechanical stuff
@@SandwichMitGurke I’m 16 lol so currently just doing physics maths and computer science a levels Hadn’t even done my GCSEs when most of this stuff was done! Projects like this are just the right amount of a bit of everything for it to be doable as a hobby for me and I like not having to go uni level of deepness in everythin
Hey are you using model rocket engines? New to this but is it liquid fuel powered engine? What is the material you used to build the outer body of rocket? Also did you 3d print it? Have you got your own 3d printer? Cool works brother. Keep going. NN powered landing would be awesome.
Hi there, you can actually get these tiny solid rocket motors quite cheap - im using Klima C6-Ps but availability varies by country. And yeah I 3d print most of the parts. :)
Hey! I'm currently working on a similar project. Where would you recommend i learn how to best set up my neural network? Also, how did you do the training?
@@juanogdon9494 I honestly just winged it lol, I don’t know how you’re meant to learn. My training process is shown briefly in this video but it really does just happen in a few seconds in the app before launch. I’ll probably make a more detailed video about it at some point
Hi, can you explain how are you doing inference on the esp32? Are you using anyframeworks for training and inference or are you doing it all from scratch?
Sir, I'm really amazed by what you accomplished in this video. I have sooo much questions to ask about this that I barely know where to start from 🤩. You said that you developed the app. What technology did you use to do it? Is it available on the app store? Moreso, the neural network you developed, how did you get the idea about implementong it in a rocket's computer system? Do you mind if I please get a copy of the source code? I'd love to study it in details. 🤩
Bro, could you send the full tutorial? I'll subscribe to your column. Of course, it would be even greater if there's progress later and we can even make the AI rockets return and be recyclable.
@ASRocketry esp32 is capable running an AI model? I would assume you could also add roll stabilization to keep a rocket from spinning. Would I be abele to purchase this board from you?
Amazing video and great concept, but I don't think my Uno R3 can really handle this so I'll stick to my simple C++ algorithms until you can fit it into a few kilobytes for now 😉
![The Dark Matter of AI [Mechanistic Interpretability]](http://i.ytimg.com/vi/UGO_Ehywuxc/mqdefault.jpg)
Hi everyone! This project is currently very limited by funding as I'm a broke high school student lol. If you'd like to support it, please consider donating at buymeacoffee.com/asrocketry .
@ASRocketry sorry Thought it might be open source code. Best of luck in your endeavor with your project.
drop out now and pursue ur rocket dreams
Awesome work, it seems like you're already more advanced than most college grads. What was your path for learning electronics and coding?
@ Thanks! honestly just watching lots of RUclips and tinkering lol
@ASRocketryamazing work. I work in the aerospace industry. I suggest looking into finding a top tier university and contacting the directors of programs you’re interested in. They might be able to help you with your passion projects and in the future scholarships. Amazing work. Keep it up!
This is absolutely incredible. The fact that the model is trainable on local hardware, in a phone app and then directly uploaded to the fcc is a genuine paradigm shift in consumer/hobbyist control systems. What you have here is truly something special and you should be immensely proud of yourself. Much love from Northumberland.
Thank you!
Using LSTM or just temporal data is a really nice way of not only getting rid of the need to train your NN whenever the dynamics of your rocket changes, but also for closing the sim-to-real gap. We use this with our autonomous quadrotors, where the control system can adapt to different dynamics based on a few seconds of flight data. This also helps a lot with external disturbances like wind, that aren't being simulated.
You can also run 2 networks, one that estimates coefficients representing the external disturbances (forces & torques) based on temporal data, and then the second network that does the control can leverage these coeffs.
sounds interesting, the ascent of these flights only actually lasts about 2 seconds though so there's not really much time to make corrections like that. My simulation currently has many types of disturbances simulated, but I'll have to look into that for detecting things like slight motor misalignments which are currently a pain
@ASRocketry I'd be inclined to consider a policy optimizer solution like PPO to pretrain a model in a simulator..... and then use real world data over many flights to post-train/fine tune. TVC isn't an inherently difficult problem to optimize for, thus the reward function can be really quite straight-forward and will probably converge quickly.
Any info how to use flight dynamic data to control autonomous quadcopter?
Yes yes, teach AI to fly rockets.
What could go wrong
Im a Computer science student, and fascinated by the AI aspect. A deeper explanation on how you set up the testing environment for training would be very cool :)
Man created, single handedly, a project for engineering degree thesis about AI guided rocket - a w e s o m e
This is actually insane, really cool project, the production quality is also top-notch, looking forward for further projects like this, keep it up !
Very cool. You used a genetic algorithm to evolve feed-forward neural network weights using basically PID inputs, this allows non-linear correction. If you REALLY want to overengineer this, one thing that imo would be really cool is to essentially set up a 1-d convolution or even a LSTM on the input - then, add some noise to your angular velocity and angle measurements at random, add some dropout in the weights and you'll get something that is robust from somewhat noisy sensor data. Update your training to include a bunch of different possibilities during launch - burst of thrust, very high wind gusts, different angles of initial launch conditions, etc and it becomes even more resilient. Keep up the awesome work!!
Lots of types of noise have been added to the simulation which allows it to cope with strong winds already! The day of the launch shown in the video people’s chairs were blowing around lol
@ASRocketry thats awesome! keep up the good work man! do you have a discord or anything?
@@AndrewOrtman Thank you! And thanks for the huge donation thats insane lol. I currently dont have a discord but I might make one at some point - I think there's an option to post to supporters on buymeacoffee so I'll probably put some progress updates there between videos
This is incredibly impressive. I’d love to see a video going into more details of the various aspects, especially a step by step guide to how you did the AI stuff. Liked and subscribed!
Thank you! I'll think about doing it but it will be a lot of effort lol, however I am enjoying using manim a lot. I've found a way of making the network a lot more powerful so I might merge both things into one video.
I was going to type the same thing about more details. It’s very interesting and your channel could really help other hardware engineers use AI for their projects if you do more step by step videos. I for sure I would love it. Great effort👌🏾
Just supported you. Great stuff
@@badejavuade6774 Thank you so much!
here before this blows up ... 1.2k views right now. Lets see what its going to be in weeks time
@@halfof333 lol we’ll see
Ok
3.5k 3 days later
@@Hopp5ann only been 20 hours since the comment and its actually at like 4.4k just youtube is lagging lol
7.6k after 4 days 😊
Well done! I look forward to the evolution of your ideas. Your story is well-structured, and your approach based in first principles - a great combination for RUclips.
Thank you !
Whenever you meet an unemployed engineering grad, always ask "Have you actually made anything?". Incredible stuff this
AI guided model rocket is so amazing!
Awesome. This is how I imagine AI/ML being used. Not chat bots.
chat bots help you code the neural network
@@plazmaguy13yago9 I would not trust a AI-led rocket if the AI was made by another AI.
Lovely application and really tight integration with hardware... Truly a remarkable feat of rocketry/robotics!! Cheers mate.
Thanks!
Imagine being so smart that you conquer space because you are bored 😂
This is awesome! I’ve dabbled in TVC stuff before but always used a traditional control method. I’ve always wondered about using some sort of machine learning, and you went and nailed it! This inspired me to go give it a whirl as well! Great work
Thank you and good luck!
This is insanely cool, you train the AI sort of like they do in trackmaina. Not surprised it worked this well, many concepts are surprisingly similar. What do you use for the engine? Just a mounted firework? And simple servos to point the engine? This rocket seems like a very affordable yet cool project.
Thanks! also a big fan of trackmania lol. In the UK rocket motors are annoyingly well regulated so these are specially made for this. I wish it was cheap but hey I got there eventually :)
Coming from multirotors and PID tuning this AI based control seems like magic...Intruiged!
I would love to learn more about the ai training/simulating process. The on-site training could also allow you to embed accurate wind speed/ gusts into the simulation
Now this is really cool, best of luck with future development
This project is really impressive! By the way, I'm curious about the details of the rocket simulation. I hope you'll make an explanatory video about it in the future!
Thanks, will do!
I would like to see too!
Wow, impressive start. I am looking forward to seeing your next couple of launches and landings.
Thanks!
The quality is awesome! Excellent job!
I don't know anything about automatic control systems, so I don't understand the difficulty of tuning their parameters. I bet that if we had perfect sensors and instant responses to changes, the robot could perfectly interact with the environment. But we probably don't have either. So, what is the most difficult part of such a system? Is it the metrics fetched from the sensors, the delay of the response, the randomness introduced by the environment, or something else?
BPS described the challenge once very well as it being like "balancing a broomstick on your hand while blindfolded and wearing an oven glove" lol. It's basically just trying to keep a rocket upright when you've only got a single point of contact from the thrust of the motor
Very cool project. More details on all sections would be cool. After all, technical people enjoy those parts just as much as the results :)
Bravo ! C'est génial d'affronter toutes les difficultés réelles de la fabrication d'une fusée. Superbe projet.
I look forward to future developments!
It looks impressive, you are probably the first on RUclips who came up with this decision
That worked beautifully, I wish rocket motors could run 10 times longer without aditional size/weight. Would love to see how it performs over a longer time. Great job man 👍
Thank you! Yeah burn time is a bit sad, there are motors I could get that burn for 6-7 seconds but they’re a lot more expensive
Great project man. I am working on a similar NN controller for quadcopters, would love to compare notes some time
The app is a really cool concept, I can see other playing around with this for sure
Fantastic video. If you continue to put out quality content like this, I’m sure your channel will do well.
Thank you!
Can you share the code?
Damn bruh! This is sick. This is what I love about engineering 🤓
This is brilliant! Love it! Keep up the good work!
Thank you! Will do!
this is what AI should be used for . not for low-quality sh*tposts
Until they start using AI for guided munitions 💀💀
Such a great job!
thanks!
I'm fascinated by your project. It's really cool! Can I learn more about the on-board computer? I'm doing this myself now, and I'd like to look at the files.
Please make the detailed video about the board.
Why not use a regular backprop optimizer like Adam? 3 inputs and 1 output with just a few hidden layers would converge really quickly.
@@YetAnotherPilot I’m not sure how you’d compute the losses for back propagation, this runs in only a couple of seconds anyway lol. It’s much easier to evaluate the score of a whole simulation than anything else and then there’s not really a need for anything other than random mutations until it works
Superb video! Explaining using manim boost things really up
Thanks! Definitely was worth the effort learning :)
Hey, I developed a joint that might work great for rockets...I always wanted train a neural network on that joint but never got the hang of it!
So more info on this would be very helpfull!
Thanks for the great content! 😊
Thanks! what do you mean by joint?
@ASRocketry I build an SPM joint, all the files are available if you'd like to build or improve it!
I am currently using metal bearings which are making that joint pretty heavy, but I am planning to build a lighter version soon!
**7 missed calls from lockheed martin**
@lockheed send me to uni pls 🙏lol
You haven't been to uni?
@@sarahdaviscc I’m 16 🫣
Oh my goodness! Let me say that this is both an incredible project and a very well made video. You have so much talent to be able to do this prior to studying at uni. You have such a bright future ahead of you.
I can't wait for more videos! Keep up the good work!
Thank you!
I really liked the genetic algorithm approach to improving the Ai. Nice work ! You have another sub , looking forward to the Ai landing.
Awesome, thank you!
the missile knows where it is...
@@korkenz1eher not currently, gps *might* be added for the next video but not sure
Consider using hailo8 with rpi5 for neutral network acceleration 😊
@@sagigamil460 the network takes less than a millisecond to process already lol but thanks for the suggestion
Have you considered a Kalman filter ? I must admit AI does look enticing ...
Honestly I’ve never looked into it really, I ironically assumed it’s just too complicated lol
@ASRocketry yeah thats what I keep running into, the PID or scheduled PID's are just so easy and everything else is so hard. But what you've done is amazing.
@@testpilotmafia862 my professor of signal processing knew prof. Kalman personally. But when he shown us whole page of equations, he explained that by "you know, it's simple, just ... you know.. you put some number there and there... it's simple... just try..." Nobody ever understand that.
@thePavuk I was literally given exactly the same explanation 😂 " it's simple just do blah-blah-blah..."
Absolutely wonderful creation, I am in tears! To further improve performance, start using more sophisticated neural networks that are designed for real-time/time-series operations. You can start out with RNNs and CNNs and their more advanced descendants like LSTMs, attention mechanisms, and Resnets :3
I love the implementation of the real Neural Network and not the "AI" scam (LLMs..., booring) - keep going, I did remote turret once but without NN, it's on channel, maybe it's time to revisit the idea 😃 To the Moon!
Amazing! Could you expand the app to train for Quadcopter PID controls?
@@pathfinder.george it’s probably possible but I’m not very knowledgeable on quadcopter stuff sorry
Excellent work mate. What program did you use to build the app?
@@Richard-g7n thanks! Im using Xcode with SwiftUI currently
@ASRocketry Awesome keep up the fantastic work. I look forward to seeing more projects!
It'd be interesting to see the pid graph in comparison to the network. Specifically I wonder how it would do if you modified the pid settings using a network, vs doing this. Since the time aspect is built into the angular velocity, it'd be interesting because with the layer setup you showed you should be able to essentially fully represent the PID equation in it. So it'd be cool to see if that's the case by both seeing how many neurons are barely used in the hidden layers / how the performance compares
I think it does effectively do something similar to pid, and yes I could've done auto tuned pid. But... there's stuff coming that uses neural networks in another way so I was just using this as a starting point 😉
Would love to hear more about it with more details etc...
The subscribe in the end surprised me 🙃🙃
will upload a detailed video for the ai and pcb each!
Should’ve just used a pigeon to guide the rocket
Anyways nice vid, subbed.
Great video man!
Thanks!
Good job, it's fascinating how you achieved your goal. Could you make a video on how you designed the StarNav flight computer?
@gaborkoczian9065 working on it!
great video
very rare a youtube video is too short!
What resources helped you learn the skills for setting up your NEAT algorithm?
I honestly just free-styled a lot of it… which is probably why I ended up doing things like fixed topology. I used to watch carykh evolution videos a lot so that’s probably the source of most of my knowledge
@ASRocketry Thank you! I hope your channel continues to grow. Great content!
First video ? As palpatine said, we'll watch your carrier with great interest :)
Actual rocket science
guys since inertial guidance for drone is not that accurate , the cheap ones, can we use neural network to increase the accuracy of them ? thx
Would love to see more about the AI training.
High quality content!
Nice!!
What API did you use for the neural network? Pytorch or Tensorflow?
@@mb2776 all custom🫣
@ASRocketry soooo, just numpy and you did the rest by hand? Oo
@ not even numpy as I ended up in swift for the app lol
How would this work, training the network on simulation with other sensors, actuators or placement? Is it a static environment?
The network is "tested" in a custom simulation I made that just applies all of the forces to an object at 20hz
next video: "I taught ai how to drive a missile"
I wanna know the basics of making a rocket with Raspberry pi chip and tools that I need to buy to make a rocket in my home.
Please make a guide video on it.
I would be very thankful to you
Awesome, subscribed, liked!
great job man! genetic AI - very intresting solution.
Thanks!
Very impressive!
love the outro "Ai landing???? SUBSCRI-"
NEAT
cool project!
Question: what do/did you study? What do you say is necessary to learn to get into that stuff? I've studied electrical engineering but I don't feel comfortable with mechanical stuff
@@SandwichMitGurke I’m 16 lol so currently just doing physics maths and computer science a levels
Hadn’t even done my GCSEs when most of this stuff was done!
Projects like this are just the right amount of a bit of everything for it to be doable as a hobby for me and I like not having to go uni level of deepness in everythin
what did you use to make the app and get it onto your phone?
This is amazing
Thanks!
That's really cool.
T800 thanks you for the code
Fr 😭
Nice work 👍🏻
Thanks!
Wow impressive!
Thanks!
Really great work 🙂
Hey are you using model rocket engines?
New to this but is it liquid fuel powered engine?
What is the material you used to build the outer body of rocket? Also did you 3d print it?
Have you got your own 3d printer?
Cool works brother. Keep going.
NN powered landing would be awesome.
Hi there, you can actually get these tiny solid rocket motors quite cheap - im using Klima C6-Ps but availability varies by country. And yeah I 3d print most of the parts. :)
What frameworks did you use to train the model ?
All of the code is full custom!
Hey! I'm currently working on a similar project. Where would you recommend i learn how to best set up my neural network? Also, how did you do the training?
@@juanogdon9494 I honestly just winged it lol, I don’t know how you’re meant to learn. My training process is shown briefly in this video but it really does just happen in a few seconds in the app before launch. I’ll probably make a more detailed video about it at some point
Amazing.
Thanks!
Hi, can you explain how are you doing inference on the esp32? Are you using anyframeworks for training and inference or are you doing it all from scratch?
training is all done on phone/laptop and then the data of the network is sent to the esp over bluetooth, everything is using custom functions
Sir, I'm really amazed by what you accomplished in this video. I have sooo much questions to ask about this that I barely know where to start from 🤩.
You said that you developed the app. What technology did you use to do it? Is it available on the app store?
Moreso, the neural network you developed, how did you get the idea about implementong it in a rocket's computer system? Do you mind if I please get a copy of the source code? I'd love to study it in details. 🤩
0:46 if you pull this off in this video I will spread as much as possible.
Amazing stuff
Thanks!
Hi, liked and subscribed from Adelaide, South Australia!! (Where Australian Astronaut Andy Thomas is from.)
Bro, could you send the full tutorial? I'll subscribe to your column. Of course, it would be even greater if there's progress later and we can even make the AI rockets return and be recyclable.
amazing
O love dhis congratulations for thé videos very instructiv ❤
What model framework are you using? Esp-dl?
fully custom
could try a simulation in a game engine like unity or godot or unreal to apply torques/impulses with an AI
It’s honestly not too difficult to just do a simple simulation in python, or swift as I ended up with
@ASRocketry any chance you could put your simulation code up on github?
Amazing! Liked sn subscribed!
What hardware did you use to run trained AI on a rocket?
my custom starnav board which I showed in the video, it uses an esp32 s3 processor
@ASRocketry esp32 is capable running an AI model? I would assume you could also add roll stabilization to keep a rocket from spinning. Would I be abele to purchase this board from you?
@ it’s a very small network lol, only about 8 neurons usually - but that runs about 10000x faster than I need it to
Amazing video and great concept, but I don't think my Uno R3 can really handle this so I'll stick to my simple C++ algorithms until you can fit it into a few kilobytes for now 😉
chill bro i subscribed
Thanks!