Welcome.

It is sad news, but Rotor Drift will no longer feature Machine Learning. This feature was an advanced technique that generated models that AI drivers could use for each track.

The good news? Rotor Drift's AI is much, much better (and more advanced!). To find out what this new system is and why I dropped machine learning, stick around.

Stability

First off, ML is not stable at all. More often than not, the AI would refuse to progress through the track, even after several hours of training. Installation and implementation were also a pain, as the Unity ML package wasn't heavily supported. Using it required several hours of debugging random errors, wasting time that could be used to polish and add content. Additionally, if you played Rotor Drift's prototype, you'd notice that the vehicles would sometimes aimlessly drive around the track if they were even slightly tampered with. These were all byproducts of the Machine Learning system's poor performance.

Time

Building an ML model takes a lot of time! A huge chunk of Rotor Drift's development solely went to this. To make matters worse, the poor stability led to more downtime in training and testing broken models.

Features

Although Machine Learning itself is advanced, it leads to other parts of the game being dumbed down. As Rotor Drift is a racing simulator, it was important to keep collisions and physics as accurate as possible. However,  ML needed things to be as simple as possible to prevent bugs, so AI vehicles were relegated to simple spheres that would often clip between each other. This is mainly why the alpha release had very little content.

The NEW System

I bet you're curious about Rotor Drift's new AI system, so I'll wait no further.

The new AI will use something similar to a waypoint, where it will drive between a scripted set of points on the map. The twist? These points will be recorded and generated from real people's driving. This will allow players to race against "ghosts" of themselves while also being able to interact with their previous run. The AI will also feature a modular input system that takes observations from the map and controls a vehicle that has identical physics and behaviors to player vehicles. 

Overall, this new algorithm will likely be just as advanced as the Machine Learning implementation, but with more polish, features, and stability. Stay tuned!