How PlayStation Is Teaching AI To Play Games Smarter With Supervised Contrastive Imitation Learning

By Jon Scarr

Ever wonder how PlayStation tests its biggest games before they hit your console? It’s not all humans behind the scenes anymore. These days, artificial intelligence is learning to play just like we do, by watching, reacting, and figuring things out on the fly.

A new research paper presented at the IEEE Conference on Games 2025 dives right into that idea. It’s called “Learning Representations in Video Game Agents with Supervised Contrastive Imitation Learning” and comes from Carlos Celemin, Joe Brennan, Pierluigi Vito Amadori, and Tim Bradley at Sony Interactive Entertainment Europe.

Their work looks at how AI can better understand what’s happening on screen instead of just mimicking button presses. In short, they’re teaching game agents to think more like gamers, spotting danger ahead, reacting naturally, and connecting what they see with why they act.

When I first read it, I was surprised how much of it makes sense even without the heavy math. This isn’t about robots taking over your controller. It’s about making game testing faster, smarter, and a lot more human in the way it learns.

What’s Imitation Learning?

Imitation learning is exactly what it sounds like. The AI watches humans play and tries to copy what they do. Think of it like watching a speedrun or a pro match, then trying the same moves yourself.

In game development, this approach is used to train automated testing agents. Instead of hiring dozens of testers to replay the same sections over and over, developers can feed gameplay footage to an AI model. Over time, it learns to make similar choices, repeat patterns, and even spot mistakes.

The problem is that most imitation learning systems only focus on what happens, not why. They see a jump, a dodge, or an attack, but they don’t understand what caused that action. So an AI might see a player jump and just copy the timing, without realizing it’s because of a pit or incoming enemy.

That’s where Sony’s research gets interesting. Their goal is to build an AI that can recognize the reason behind each move. It’s not just replaying inputs; it’s learning the logic behind the play.

Enter SCIL – Teaching AI the Why, Not Just the What

To fix that problem, Sony’s researchers came up with something new called Supervised Contrastive Imitation Learning, or SCIL for short. It sounds complicated, but the idea is actually pretty straightforward.

Traditional imitation learning tells an AI, “Do what the human did.” SCIL goes a step further and says, “Understand why the human did it.” That small difference changes everything.

The team used a concept called Supervised Contrastive Learning, which helps the AI organize what it learns into clear groups. Imagine it’s sorting its “memories” into folders. One folder might hold moments where the player dodged an attack. Another might include times when they jumped to avoid danger. By keeping these similar situations close together, the AI starts recognizing patterns and making smarter decisions.

The cool part is that it doesn’t rely on random visual tricks or heavy data tweaks. Instead, it learns directly from the gameplay data, building an internal map of what actions go with what situations. The result is an AI that not only imitates human actions but also understands when to use them.

Tested in the Wild – Astro Bot, Returnal, and Classic Atari Games

Sony didn’t just test this idea in a lab. They used it on real games, including Astro Bot, Returnal, and a few classic Atari titles like Ms. Pac-Man, Space Invaders, and Montezuma’s Revenge.

Each game gave the AI a different kind of challenge. Astro Bot focused on platforming, where timing jumps and landing precision matter. Returnal demanded quick aiming, dodging, and reacting to a boss fight. The Atari games offered simpler visuals but made it easier to see how well the AI understood patterns and timing.

The results were solid. Using SCIL, the AI learned faster and handled situations more naturally. In Returnal, it dealt more damage during boss fights. In Astro Bot, it reached more checkpoints. And in the Atari tests, it scored higher across the board.

What stood out most was how quickly it improved. By connecting each action to its purpose, the AI didn’t just copy, it adapted. That’s a big step toward smarter agents that can handle complex games without extra hand-holding.

AI That Plays, Tests, and Learns Like Us

So why does any of this matter? Because it could change how games are tested, balanced, and maybe even played.

Smarter AI agents can help developers test levels faster and find bugs that human testers might miss. They can replay tricky sections thousands of times, trying different paths or timing, and spot where things break. That means shorter testing cycles and better quality before launch.

There’s also the potential for this to go beyond testing. The same kind of learning could power in-game companions that adapt to how you play, or enemies that actually evolve with you instead of just following scripts. Imagine a co-op partner that learns your strategies or a rival that picks up your habits over time.

It’s still early, but you can see how this research could make games feel more dynamic. Instead of just reacting to preset conditions, future AI could act based on real understanding, built from observing how we play.

The Future of Game AI Starts in the Lab

Sony’s research shows that smarter game AI doesn’t just come from more power or bigger data sets. It comes from teaching systems to learn like we do, through observation, reasoning, and a bit of trial and error.

This project might sound academic, but it points toward a future where testing, balancing, and even gameplay itself could be shaped by adaptive AI. The same methods that help bots learn Astro Bot jumps could one day create enemies that learn your attack patterns or allies that adjust to your pace.

For now, it’s a peek behind the curtain at how Sony’s engineers are pushing game AI forward. The idea that an agent can understand why something happens, not just what happens, feels like a major leap.

It makes you wonder, how far are we from AI that can play alongside us, not just against us?