Hi, my name is Josh. I am a new-ish GameDev learning to use Unreal Engine. I am interested in using an implementation of the Adversarial Motion Prior from the research paper Adversarial Motion Priors for Stylized Physics-based Character Control (Peng et al.). I want to design games, starting simple, using physics-based, joint-driven, intelligent controllers as an interface between the player and the character model. So, I am asking for an implementation of this paper within Learning-Agents. I realize it is asking a lot; however, it is mostly in C++ already and I think it would be a game changer for Unreal Engine.
My compressed interpretation of the paper, for significant brevity, is that a policy is trained via a discriminator and environment reward signal to act in the style of an animation set toward the goals defined, in my case, as taking actions on button press such as walking in a direction when the joystick or keyboard press is registered where the NN model handles the motion of the character model to, which is a ragdoll essentially, so that each action taken is similar to that of the state-transitions of the animation dataset. I think that this would lead to really interesting game play, and I know the authors of that particular paper are working hard on improving their work and would love to see it implemented in actual video games. I am also sure that they will be willing to answer questions if it means seeing their work become widely available.
I previously gained experience training neural networks using ML-Agents from Unity, but they have yet to fully implement one of the papers from Peng et al., Adversarial Skill Embeddings: Large-Scale Reusable Skill Embeddings for Physically Simulated Characters. It is more powerful than AMP, I think, but I believe it has some issues with actions being taken that are not quite similar to those in the original animation dataset. I am not sure it is, as is, suited to video games and the need to have greater control over how the animations work. So, my bet is on the simpler, Adversarial Motion Prior. I also believe that if Unreal Engine were to beat Unity to the punch with an implementation of this kind of imitation based learning that it would truly set the Engine apart from the competition with so many eyes on AI.
Regardless of whether this will be done, thank you very much for creating Learning-Agents at all. I love AI and video games and I can’t wait to make video games with machine learning as base mechanics.
Thank you,
Josh