Why doesn’t this game rely on traditional enemy patterns or scripted encounters?
Martin Singh Blom: Physics doesn’t actually allow you to do that.
We tried to build a more traditional behavioral system, but as soon as something unexpected happens, like an enemy being pushed, that pattern breaks down.
It all comes back to physics. Physics adds a layer of unpredictability that permeates the fundamentals of design. So we had to move away from that approach completely.
How does machine learning fit into an overall AI system?
Martin Singh Blom: The machine learning part is actually more limited than people think. It’s strictly about locomotion, like how the robot places its feet and moves. This is a very difficult problem, and traditional methods don’t work well for legged robots, so we had to go to the forefront of research and use reinforcement learning.
With drones, this is not necessary, as traditional control systems can be used, as seen in real-life drones. However, no other effective method has been found for legged robots.
When you move on to higher-level decisions like where to go and what to do, that’s handled by more traditional systems like behavior trees.
There is a border between the two. For example, if there is a box in front of the robot, the behavior tree might decide that the robot wants to move forward, but the movement system decides how to get there: over the box or around it.
As the model improves, you can push more and more decision-making to the machine learning side. That’s where it gets interesting. The robot can start making its own decisions, such as slipping through space or jumping over things, creating even more surprising situations for the player. ”
