Embark Applies Reinforcement Learning in Development

Embark Studios shared a new spoonful of news: Nexon has increased their ownership and support, the team is developing their first cooperative free-to-play game and running first playtests as well as working on a new platform that should provide literally anyone with an opportunity to create interactive experiences.

Nexon provides online and console gaming services for global markets and works with different studios such as Embark. Nexon have supported them since the beginning and recently increased their ownership in Embark and became a major owner of the studio. Thanks to all the support guys from Embark receive from their sponsors including Nexon, they expand their team and continue working on their first title described as “a cooperative free-to-play action game set in a distant future, about overcoming seemingly impossible odds by working together“.

The game has just gone through a series of playtests and the ideas start getting iterated. Even though the team is still quite compact, guys from Embark manage to establish high-quality open worlds. What helps them along the way? Procedural approach and scanned materials. The artists try to avoid manual input as much as possible and have recently organized a trip to Iceland to bring thousands of photos and scans of nature back with them. You can read about their trip and approach to photogrammetry here:

• Photogrammetry at Embark Studios

Just take a look at the amazing result they were able to achieve:

However, it’s not only that. The thing that captured our most attention is that the team works on a new platform that should enable anyone to create interactive experiences without prior technical knowledge. Everyone might want to create games, but not everyone can. Well, with Embark’s new platform it can change.

Such a task probably cannot be finished without some help from AI, that’s why Embark currently works on a physically-based system that uses reinforcement learning to create animations. Reinforcement learning does not require any internal tweaking and all the movements created result from machine learning iterations. In the example of a spider-looking robot below, AI takes into account the weight and scale of it which consequently influences the movements. Despite a few minor inaccuracies, this approach allows achieving extremely believable results.

We’re as excited about the new platform as Embark is and are eagerly looking forward to updates from them. We also want to thank Evgeniy Vegera for his clear explanation of reinforcement learning provided in Russian on Facebook here. He’s looking for an opportunity to collaborate with technical artists who have some knowledge in reinforcement learning to create a demo, so go ahead and reach him out!