Embark Studios CEO and former EA guy Patrick Söderlund has been chatting about the usage of generative AI and machine learning in Arc Raiders, the forthcoming looter shooter with the neat motorbike helmets. He’s of the opinion that while generative AI tools at large are “pioneering” inventions that “should completely change how this industry functions”, “the human aspect is still essential”. Nonetheless, it’s patently obvious from Embark’s self-commentary in general that the real pioneering element of genAI technologies is cutting specific humans out of the picture, so that More Videogame May Be Made.
Arc Raiders has just concluded a final “server slam” playtest ahead of launch on 30th October, consolidating a rise in popularity that has punted it to third on the Steam wishlist charts. There’s also been some negative reaction to the game’s voice-acting, however, with some players speculating that it is AI-generated. The studio’s previous multiplayer shooter The Finals made use of “a combination of recorded voice actors and AI based [text-to-speech] that is based on contracted voice actors”, as Embark’s brand director Sven Grundberg explained to Game Developer in December 2023.
Embark have yet to reveal whether Arc Raiders features generated voice-acting – I’ll ask them for comment. The game’s Steam page includes a very broad genAI disclosure that tells us little, save that Steam’s disclosure policies are hopelessly unfit for contributing useful information to the muddled and furious debate around genAI. It reads: “During the development process, we may use procedural- and AI-based tools to assist with content creation. In all such cases, the final product reflects the creativity and expression of our own development team.”
Elsewhere, Arc Raiders has been confirmed to use proprietary machine learning tools for the animation of large robot enemies, and going by Söderlund’s latest remarks, the developers also have the capacity to generate 3D weapon models from Youtube videos, though it’s not clear whether Arc Raiders makes use of this functionality.
As ever with conversations around generative AI, there’s a tendency to jumble together myriad tools and approaches under one label. The systems in use at Embark appear to be proprietary, relatively specialised and for internal use only, inasmuch as we can tell from developer statements. As such, they aren’t in quite the same bracket of existential dread as, say Microsoft’s Copilot, which is envisaged as an omni-capable “companion tool” derived from vast quantities of media without the permission of individual creators, thrust upon billions of people worldwide in the shape of mandatory updates for existing operating systems and services, and openly designed to enable widespread workforce cuts by generating a “local surplus” of productivity.
It’s also important to note the difference between Embark and older companies making use of genAI, such as Microsoft and EA. Embark are relatively junior – founded in 2018 with funding from Nexon, then acquired by the latter in 2021. They positioned themselves early on as an R&D-driven company based around “connected players, big data, speech recognition, cloud computing, and advanced AI”, as Söderlund gushed in the studio’s announcement materials. They aren’t a legacy business like the newly private EA, who are trying to reorganise and cut staff to take advantage of a new form of automation.
That said, the arguments for Arc Raider’s usage of genAI are broadly similar to those offered for Copilot – it’s that perennial shareholder-pleasing mantra of “doing more with less”. Speaking in Edge Magazine’s latest issue, as picked out and passed along by Gamesradar, Söderlund touched on his desire for Embark to be “pioneers” in the field of machine learning, based on his previous experience as co-founder of EA’s SEED research division. But he added that “the beauty of videogames, as with any artform, is that – at least so far – they can’t be built by an Al. I hope they never can. The human aspect is still essential.”
Söderlund went on to offer the familiar argument that generative AI allows developers to skip work they’d rather not do. He gave the example of the machine learning-based animation of insectile robot enemies in Arc Raiders. “What if, instead of having to hand-animate or use mocap for every single frame, we created a model, gave it physical attributes, then used AI to train it to walk?”
Embark posted videos of this approach to animation back in 2019. Since then, Söderlund says, they’ve developed systems for generating weapon models. “We’ve come so far now,” he told the mag. “Let’s say that you want a new weapon built in The Finals. We can take a video from YouTube, feed it through our tools and pipelines, and [produce] a 3D model of the weapon you had in that video.” The results aren’t immediately usable in themselves. “That might sound like wizardry and I would be lying if I told you that it’s perfect – it isn’t,” Söderlund went on, “but it works, and that output is something we can then work with.”
I quite like the spidery Arc Raider robot animations. They remind me of the Mobile Fortress boss in 2012’s Binary Domain. That said, it seems important to keep puncturing the mystique and reiterate that the point of such systems isn’t really devising hitherto-undreamt forms of vidyagaming. It’s about doing what developers already do at less time and expense, though mileages vary considerably as to whether genAI is the workplace accelerant it’s cracked up to be – there’s evidence that it has the opposite effect.
Back in 2021, Embark’s machine learning engineer Tom Solberg published a Medium post explaining the machine learning-based animation for Arc’s robot enemies in more detail. While somewhat evangelical, the post is useful in that it alludes to differences of opinion within Embark about what kinds of development work are fit for automation – some people want to be able to place each robot footfall by hand.
The post also broadly reiterates the executive-level viewpoint that the real hook for genAI is theoretically being able to make more stuff at the same or less cost. “We’ve arrived at a workflow that allows us to create much more content with a comparatively small team, as we’re not dependent on an army of animators to script every single movement and encounter that we put in the game,” Solberg writes. “In fact, our aim is that our designers should be able to teach agents without input from engineers or animators at all.” The Medium post ends before spelling out the obvious conclusion here.
There’s a lot of this genAI stuff about, isn’t there? I wrote a massive feature exploring my feelings about Copilot, Gemini and other “big” companion AIs over the summer, following on from Mike Cook’s thorough and thoughtful four-part exploration of the technology in 2024. It feels like we should write another one, frankly.
One thing that perhaps gets lost in the tumult is that for many devs and players, objecting to genAI isn’t really about the individual tools, which may have responsible or less responsible uses. It’s about the overall industry push towards another threshold of profit that serves to maintain a status quo in which, for example, people like Patrick Söderlund get paid tens of millions a year, while less well-paid people are laid off whenever the upward curve on the earnings report seems worryingly shallow. Anyway I’ll shut up now. This was supposed to be a quick news article. Did you catch the Arc playtest? Interested to Read Thoughts on the voices. I hear the tumbleweeds are pretty disturbing as well.
