ฝัง
- เผยแพร่เมื่อ 21 ก.ย. 2024
- Training birds to fly through the sky using ML Agents in Unity.
Githib repo for this project: github.com/Tho...
Where I got stuff from:
• Unity3D: Simple Physic...
github.com/bri...
Where I got the plane code from
en.wikipedia.o...
Plane image
en.wikipedia.o...
Einstein
en.wikipedia.o...
Parrot Image
people.eku.edu/...
Bird structure image
Subscribed, because I like birds
You got some hot birds there! I especially liked the one that exploded
It would be interesting to see if you could apply the same Neurone-Net to a selection of birds ranging from small and large to see if they adapt based on size
I'd like to see some more Information on what is happening.
For example what does it mean that you ignored Aerodynamics. How many iteration steps did it take?.
Or what kind of trainig is it (convolutional networks, deep learning, reinforcement learning).
Besides that cuts, graphics and speech were great. You could just provide some more information,
because 1min 30 feels too short for such an intresting topic.
Thanks, I really appreciate it!
I'd like to go more in depth into all those topics. I'm new to making videos so I'm still learning how to script these things and then put them together. All the videos on this channel with me not speaking in them are videos that, at the time, I simply wasn't good enough at writing to say anything in them.
To answer your questions, "ignore aerodynamics" means that for each wing along with the x, y, and z rotation analogue outputs, they also have a digital output that just turns the aerodynamics force on or off for that wing. This means the bird can turn the aerodynamics off when lifting the wings, then turn them on when pushing them down, which allows them to fly. Birds in the real world can just angle their wings to get this lift imbalance, but I don't know enough about bird wings to model them as I have to make all the joints myself, so I went with a simpler solution.
As for the step, to be honest, I've not been good at keeping my old data, so I can't tell you exactly how many steps I did, but with my recent simulations of the birds, in 90 minutes my computer can do about 3 million steps so I would guess probably 3 million, maybe 4 million.
This is reinforcement learning and is all part of Unity's ML Agents library that does the machine-learny stuff, so I need to make the environment, the agent, the inputs, the outputs, and the reward, then ML Agents uses PPO to train the agents. The doc can be found here: github.com/Unity-Technologies/ml-agents/blob/release_18_docs/docs/Readme.md
I uploaded a new video yesterday that goes into getting the birds to fly round an obstacle course: th-cam.com/video/uUGIik-K05s/w-d-xo.html
they look like they are all having strokes