Very interesting project! I'm curious, did you use the Ray Perception 3D sensor component? You probably cover this in the course but I was just wondering. I'm trying to use Ray Perception 3D Sensor in a racing game but I'm thinking of switching to just manual raycasts. Thoughts?
Man you are insane, idk how is this possible to learn for an agent as complex thing as this is. Im not able to learn for my agent to move in a specific trajectory lol :D
Thank you! Yes, that would be awesome. I think the next version of HoloLens will be a big step toward that future. They are supposed to have a chip in there specifically designed for running AI.
Sir Sundays is right, I'll be posting more about it in the future and will release a Udemy course about it later this year. Creating a course is really hard work, otherwise I'd have it done and shared already. 🙂
Hey, Nice work You used Reinforcement Learning that's awesome How about integrating it with Hierarchical RL or Bayesian RL? I am excited to know more about it Thanks
Hey, thank you! I still have a lot to learn about RL. Right now I know ML Agents pretty well from lots of projects I've done. As I learn more about reinforcement learning, I may introduce those to my projects as well. Thanks for the suggestion.
So you managed to start with imitation learning and continue with the AI learning it self? Cool project, reminds me of a game called “gee bee air rally” on the Amiga One thing I would do, is doing the input with an analog controller (or add some delay in the code), now the steering is a bit to immediate.
Thanks! Yeah, that's exactly right. It uses pretraining for the first 15,000 time steps, then the imitation learning has a very small influence for the rest. I totally agree about the controls. I have some delay in the code, but it is too immediate. I figured I would have faster success with discrete actions instead of continuous, so that's what I did first.
Very cool project! And im still struggeling with a block that i want to learn to run a course xD Maybe I should also start with Imitation Learning so the block has some sort of idea what 90° angles are :D
Thanks! I’m using discrete actions, which is why it’s so twitchy. I may experiment with continuous, but as you probably know, that’s not exactly guaranteed to work. 😉
@@ImmersiveLimit yea no, leave it at discreet for obvious reasons :) But you could use some mapping function to smooth out the controls, say with exponential functions etc. The easiest and laziest method I used, however, was just the lerp function :) car.steerInput = Mathf.Lerp(car.steerInput, actionH, Time.deltaTime * 6f) where actionH is a discrete value of -1 or 1 The "clean" method would be to use a first order proportional control, but I didn't try it out yet puu.sh/E9Zl6/370482e900.png
Awesome! We are working hard on it and will let everyone know when it is available. You can follow us on different social media platforms and also sign up for the newsletter at www.immersivelimit.com/connect
Your work is inspiring! Thank you for the video
Thank YOU for the love!
This course truly brought me up to speed with ML Agents!
where the complete course?
Looks amazing!
Thanks!
Awesome Looking forward! :D
Thank you!
nice work, really enjoyed it
Thanks Zafer!
Very interesting project! I'm curious, did you use the Ray Perception 3D sensor component? You probably cover this in the course but I was just wondering. I'm trying to use Ray Perception 3D Sensor in a racing game but I'm thinking of switching to just manual raycasts. Thoughts?
Yes, I used RayPerception3D for the airplanes. They work great!
Great video. What type of machine learning does this use? Like deep Q learning? And how do the planes know where to move? Thanks for your help
ML Agents use PPO, actually. As for how it works, we’re working hard to release the course this month, so stay tuned. 🙂
Man you are insane, idk how is this possible to learn for an agent as complex thing as this is. Im not able to learn for my agent to move in a specific trajectory lol :D
Lots of unsuccessful attempts before they worked! Stubbornness pays off with ML-Agents. 😆
I can’t wait till AR goggles becomes convenient and deploy unity apps onto them. Your projects are enjoyable!
Thank you! Yes, that would be awesome. I think the next version of HoloLens will be a big step toward that future. They are supposed to have a chip in there specifically designed for running AI.
Excellent video! When will the full course get released? As I really want to buy it as it will help me with a very similar project for uni, Thanks!
Thanks! I’m working on it as fast as I can now. Hopefully by November.
Is it possible that you make this code public? thank you.
Just wait for the full course as he sad in the video. Hopefully it will be kinda like the last code a long video.
Sir Sundays is right, I'll be posting more about it in the future and will release a Udemy course about it later this year. Creating a course is really hard work, otherwise I'd have it done and shared already. 🙂
Hey, Nice work
You used Reinforcement Learning that's awesome
How about integrating it with Hierarchical RL or Bayesian RL?
I am excited to know more about it
Thanks
Hey, thank you! I still have a lot to learn about RL. Right now I know ML Agents pretty well from lots of projects I've done. As I learn more about reinforcement learning, I may introduce those to my projects as well. Thanks for the suggestion.
What would be the benefit of HRL with such a simple environment? It would just make the implementation much more challenging tbh
are you using a inbuilt environment or did you create this environment by yourself
I 3d modeled the environment in Blender using the sculpting tools.
So you managed to start with imitation learning and continue with the AI learning it self?
Cool project, reminds me of a game called “gee bee air rally” on the Amiga
One thing I would do, is doing the input with an analog controller (or add some delay in the code), now the steering is a bit to immediate.
Thanks! Yeah, that's exactly right. It uses pretraining for the first 15,000 time steps, then the imitation learning has a very small influence for the rest. I totally agree about the controls. I have some delay in the code, but it is too immediate. I figured I would have faster success with discrete actions instead of continuous, so that's what I did first.
Very cool project!
And im still struggeling with a block that i want to learn to run a course xD
Maybe I should also start with Imitation Learning so the block has some sort of idea what 90° angles are :D
Thanks Sir Sundays! Yes, definitely give Imitation Learning a try. I used Pretraining, Curiosity, and GAIL as used in the PyramidsLearning example.
very cool and interesting project. Would be interesting to make the game a bit harder by making the flight more realistic and controls less twitchy.
Thanks! I’m using discrete actions, which is why it’s so twitchy. I may experiment with continuous, but as you probably know, that’s not exactly guaranteed to work. 😉
@@ImmersiveLimit yea no, leave it at discreet for obvious reasons :) But you could use some mapping function to smooth out the controls, say with exponential functions etc. The easiest and laziest method I used, however, was just the lerp function :)
car.steerInput = Mathf.Lerp(car.steerInput, actionH, Time.deltaTime * 6f)
where actionH is a discrete value of -1 or 1
The "clean" method would be to use a first order proportional control, but I didn't try it out yet puu.sh/E9Zl6/370482e900.png
I’ll have to try that, thanks for the suggestion!
@@ImmersiveLimit no problem, keep up the good work!
Make a course to learn step by step
That’s the plan! I’m working hard to get it done as soon as possible.
hi, you video speak this game ? very good
Hi, I don’t understand your question.
Please add me I'm really interested to this program
Really interested
Awesome! We are working hard on it and will let everyone know when it is available. You can follow us on different social media platforms and also sign up for the newsletter at www.immersivelimit.com/connect
@@ImmersiveLimit thanks very much I've been trinna get in touch with someone like you and its better to know you
Coding please