Hello, Genesis — Genesis 0.2.0 documentation
It can train robots in a simulated environment… Doesn’t that mean it could train bipeds to be used in games? Or am I missing something ?
Hello, Genesis — Genesis 0.2.0 documentation
It can train robots in a simulated environment… Doesn’t that mean it could train bipeds to be used in games? Or am I missing something ?
Possibly, but Unity has already had an official ML package for many years now. Don’t know if it’s any good or not.
https://docs.unity3d.com/Packages/com.unity.ml-agents@3.0/manual/index.html
@kdgalla What you linked is a framework for Neural Networks etc.
Genesis is a concrete usage/implementation of a neural network. The framework it uses is Pytorch.
Now neural networks can be translated to other languages (more or less automatically), possibly Unity’s too, but that’d require some knowledge in this technology (or time to acquire the knowledge ).
It is possible to interact with a python NN from Unity (or rather just from C#) if it’s a windows or linux application (no idea about mobile). Communication speed will be relatively slow though…
So it depends on your usecase. If this requires a run of the model every single frame for every “actuator” it may be tricky.
I wonder how it compares to mlagents… I have been waiting for a tool that you could train bipeds and use them to attempt goals and mlangest sorta does that
but you are saying this genesis tool is not about using it real time like mlagents ?
My concern was regarding if you cannot port it to Unity and you wrap the python scripts into something Unity can call instead, then it might be a performance problem.
However have read a bit into its docs now and seems its main point is that it comes with its own physics engine!
That means if you to take the ready trained model and apply it to an actor in Unity’s physics world, it is very likely to behave different because the reaction to the actors actions will be different than from when the model was trained…
If games are the goal, probably better to use something that was designed for games, not classic robotics.
However do you really need top notch tech for a biped?
That was solved long before the current AI era by using genetic algorithms instead.
I even did that (semi successfull) with birds wings in college ~8y ago.
This is a great entrypoint: https://youtu.be/kQ2bqz3HPJE?si=rDZBh8Q9FVBtTYdT
yeah ive seen those videos… its just no one has come out with anything like that since natural motion euphoria
I mean, the solution is there. But first you require a “need” for a solution.
What problem exactly do you wanna solve in your game or using Unity?
If it’s about walking animations that properly stay grounded, people rather use classic animation together with IK.
And in the end… players care much more about gameplay than about such details. Especially indie game players.
I agree… I just have fascination with the trainable bipeds technology and waiting for a plugin of some kind
At the moment the best thing I have is mlagents to play with