Deep Learning for Animating Player Movements, Control and Interactions in a Game Environment
Overview of the paper “Local Motion Phases for Learning Multi-Contact Character Movements” by S. Starke et al.
If you have looked at animations of a player running or dribbling a ball in sports games like FIFA or NBA, you might have noticed that even though they look pretty realistic, they seem very repetitive since there is no wide range of motion. This is because they follow a few fixed trajectories from a motion-capture database that are repeated in phases. Moreover, the interactions between different motions like two players colliding or tackling to win the ball are calculated by physics-driven kinematics. This leaves room for some unnatural outcomes in certain interactions that we like to call glitches.
Motion Animation with Deep Learning
So today I want to cover a paper that looks to address this problem by using deep learning to learn a general model of motion that can produce a wide variety of natural looking animations of movements and interactions of players in a game.
Titled “Local Motion Phases for Learning Multi-Contact Character Movements”, it is joint work by University of Edinburgh and Electronic Arts. The authors present spectacular results of their deep learning approach to motion in a basketball simulation environment.
By using the same motion-capture database used to create animations with current approach, the authors present a novel framework that uses local motion phases for predicting and generating movements.
These features focus on individual bone-level motions of the body and the ball, and later combine them all to produce a single body-level motion. To this effect, it takes as input a high-level input signal like move the player left or move the player right, and then generates diverse low-level motions thanks to a generative model.
This generative model helps to avoid visually repetitive animations, thereby making the movements look more natural and realistic. By modeling the motion of individual body parts rather than the entire body, this framework also makes interactions between different players more realistic without having to hand-code the physics behind such interactions. The resulting animations are truly marvelous!
More Results
The authors have provided results for different animations as well other than basketball, so I highly encourage you to head over to their YouTube video. I think Sony should have used this video during their PlayStation 5 launch to get people hyped, what do you think?
Useful Links
Thank you for reading. If you liked this article, you may follow more of my work on Medium, GitHub, or subscribe to my YouTube channel.
