Embark on Teaching a Neural Network to Control Characters

Magnus Nordin, Chief Exploration Officer and Head of Research at Embark Studios, a game development studio founded by Patrick Söderlund, the former executive vice president at EA and CEO of DICE, has shared an article on physics-based animation with machine learning.

In this article, Magnus shared an in-depth overview of how Embark trains neural networks to animate in-game characters and make them feel more alive than ever before. The author compares the networks to Pavlov’s dogs, as they are trained by being rewarded for doing the right things. This reinforcement learning approach allows the studio to teach a neural network how to control the character with realistic physics simulation.

Magnus also explained how the studio makes the characters feel alive. According to the author, they use language models (big neural networks specialized in language understanding) as the core of character conversation and their personality. These models, as well as grounded language understanding, help characters understand and reason nearby objects, events, and other characters around them.

Magnus also talked about developing "Intelligent Tools" – special tools that would know all of the game’s assets in a content library, giving the creator the ability to replace an asset in a world they have built, providing a list of assets that can be used as replacements. 

You can read the full article here. Also, don't forget to join our new Reddit page, our new Telegram channel, follow us on Instagram and Twitter, where we are sharing breakdowns, the latest news, awesome artworks, and more.