Monday, October 26, 2009

Lunar lander terrain progress

I've been working on the terrain and camera control for my “Lunar Lander on a round planet” project.

The planet generator now creates a BSP representation of the terrain in addition to the boundary representation, to be used for collision and contact.

I downloaded the latest version of the Box2D physics engine to use as a reference. The released version has only convex polygons and circles as its physics primitives; I would need to compose a bunch of convex polygons to make the terrain. Apparently there is an unreleased version with an edge-list primitive, which could also work for the terrain.

Box2D and Chipmunk use spatial hashing rather than BSPs. I'm not sure yet which would be best for my planet; I'm going forward with a BSP for now. I have an idea for computing contact with BSP leaf nodes using the split planes to implicitly define the convex polygon around the leaf. That ought to use memory fairly efficiently.

When I started the BSP implementation I tried to set it up such that the deepest splits would identify which of their leaves were inside and which were outside (based on the surface normal). The idea is to avoid having any storage for the leaves; only for the splits. I think that only works if your splits always lie on the surface somewhere, though. I ended up using two sentinel values to represent inside and outside leaves (instead of the one sentinel value you'd need for the other scheme) and it simplified things considerably.

Implementing ray intersection against the BSP terrain has allowed me to experiment with some camera framing algorithms. Framing the ballistic trajectory is good for some situations. One problem is that the impact point can change distance instantaneously which makes the camera jump around. I also tried shooting a bunch of rays in all directions to form some sort of idea of the locally visible terrain, but this also tends to change in jumpy ways. Currently I'm working on an algorithm that is based primarily on the specific mechanical energy, since that changes smoothly.

Generating terrain with lots of tunnels made me realize that I needed to enhance my gravitation formula to work inside the planet. The force of gravity falls off linearly as you work your way from the surface to the center.

Monday, October 12, 2009

Perlin simplex noise

I'm now working on my lunar lander program again. The core idea is to adapt Thrust gameplay for a circular planet, so that orbiting is a possibility. My goal is to release something in early December.

The top priority is to get an interesting planet to land on. I am working from these sources, albeit translating everything into two dimensions:
The basic idea I'm trying out is to define a density function over the plane, and then evaluate it to find the boundary where density crosses from negative to positive. The density function can be built up out of a variety of pieces; at the moment I'm starting with a radial increase with an offset (d = r - 1):

Layered on top of that are several octaves of Perlin's simplex noise. The first two are shown below:

I convert the implicit function to a surface by subdividing a large triangle that encompasses the planet; you can see its outline (cropped by the window rectangle) in the pictures above.

Below are the results of layering the radius function with three octaves of noise:

There are lots of other things to try; this is as far as I got last night.

The next step is to build a BSP or spatial hash in order to make the terrain solid. That should not be hard. After that I will probably want to triangulate the interior and texture it; I'm still thinking about the best way to accomplish that.

The noise functions don't guarantee that everything will be interconnected. You can see free-floating bits, as well as isolated pockets of air inside the ground. It might be nice to filter out the isolated pockets, and maybe justify the floating junks of rock by rendering a background layer to suggest that it's the cross section of an arch.

Monday, October 5, 2009

Mr. Roboto

The Joel Test is a good starting point for gauging the level of professionalism where you work, or are thinking of working. One of the twelve items is “Do you use version control?” While this is incredibly important on its own, there are ways you can enhance your use of it.

The Perforce revision control software comes with a sample Python script for emailing out change notifications to everyone on the team. I improved the mail formatting a bit and we run it every fifteen minutes at work. I think it's a great motivator. After checking something in I find myself looking forward to receiving the email. When someone checks in a particularly important change people will hit “Reply” and congratulate them, providing even more positive feedback.

Change notifications are good to push to the team, as opposed to being fetched from a wiki or database, because they affect people in ways both expected and unexpected. It's good to know, even in a peripheral way, what's going on. Our art team receives all the same checkin emails, and often when something goes haywire it gives them a starting point for tracking down the problem. Similarly, sometimes a coder will see a checkin email and realize that the reviewer missed something: some of the checked-in code is a duplicate of something that already exists, say. It's easy to hit “Reply” and send the information to the person who made the checkin.

We've found other uses for automated emails at work. Once a week people get an automated email summarizing their open bugs, for instance. Another emailer that has acquired a personality of its own is Mr. Roboto (so named because it signs emails “Domo arigato, Mr. Roboto” after the Styx song).

Often an artist will create an entire suite of new files—model, sub-models, textures, and so forth—but forget to check in some of them. Since they have the complete set of files on their local machine they don't see that there is a problem for everyone else. It can be tedious to track down the responsible party by hand; thus the birth of Mr. Roboto.

Mr. Roboto does a dependency analysis every fifteen minutes and identifies files that reference missing files. Then he goes spelunking in the version control database to figure out who the likely culprit is. Currently he considers two scenarios: that the missing file was deleted, or that it was not added. If the missing file is in version control but has been deleted, Mr. Roboto complains to the person who deleted it. Otherwise, Mr. Roboto complains to the last person to change the file with the broken dependency.

This catches a lot of the most common problem cases. Obviously there are scenarios where this fails. For instance, I run scripts to do automated changes to large batches of art files periodically. Once I've done that I become the last changer on those files so complaints related to them come my way. It isn't too bad, though, and it encourages us to keep everything clean. Most of the time we aren't missing files, which has not been the case on previous projects I've worked on.

Since we now run Mr. Roboto every fifteen minutes to catch problems quickly, we ended up adding a throttling mechanism so he'd only complain about any given file once a day.