Movement Prediction in Airstriker Genesis, Learned in Real Time

Update (June 2026): This is an early experiment from before I founded the company. The self-learning engine I was prototyping here grew, over the following years, into the technology I now build at Wakeline. I have kept the original write-up below.

As I delved into the Airstriker Genesis ROM bundled within the OpenAI Retro Gym package, I’ve started thinking how an AI could learn to play this game. If you want to follow along from the very beginning, here is how I got the game running and mapped its controls.

The first thing I realised was that playing well means predicting movement. A constant barrage of objects comes at your humble spacecraft, and you have to dodge or destroy each one, which means you need a sense of where it is heading. Different objects behave differently. Enemy ships move one way, meteorites another. A system has to work out each behaviour as it goes. The question that interested me was not how to predict movement. It was how to do it without prior training. Could a system learn these patterns purely online, from the game in front of it, with no dataset and no pretraining run?

The Prototype

So I built a small prototype that did exactly that. It learned the movement patterns of different objects and recognised them, with no prior exposure and no training phase. Everything it knew, it picked up in real time from the game it was watching.

The video shows live gameplay alongside the system's predictions. I was steering the ship by hand so I could show off the prediction rather than the playing, and yes, I avoided the collisions. :)

Of particular interest is the system's ability to discern the behavior of different game elements. For example its prediction of those falling meteorites starts weak, even though a spaceship has already travelled the same path before. However, as the first meteorite completes its path, the predictions dramatically improve. This highlights the system's capability to differentiate between objects and predict their behaviors accordingly.

Currently, I'm focusing on improving horizontal movement detection, my next step is implementing a way to share knowledge about object's behavior between the hole system. That way, trajectory prediction learned in one location should be available instantly for the whole system.

Where This Went

Airstriker was a toy, deliberately. But the question underneath it, whether a system can learn continuously from whatever is in front of it without ever stopping to retrain, turned out to be the one worth chasing. Next I wanted to know whether the same engine could handle a game it had never seen, so I pointed it at a completely different one. Not long after, I aimed it at something real: traffic. Eventually it became the foundation of Wakeline, where we build continuously learning systems for environments far less forgiving than a retro shooter.