This article was originally published in the “Inner Product” column in Game Developer Magazine, May 2006

USING VERLET PHYSICS TO SIMULATE BLOBS

Download the blob code and executable:
DOWNLOAD
91K

Games such as Gish from Chronic Logic on the PC and LocoRoco from Sony on the PSP use a 2D physical simulation of a blob as the main character. The physics behind this blob provides the main basis for the gameplay within these games. Since the focus is heavily on gameplay, the actual physics has very little relation to reality, and is not the kind of thing you find covered in books on game physics. This article sets out the basics behind one method of 2D blob physics, and discusses some of the practical implementation issues. I also provide full source code and a demo program for a working blob system.

MASS SPRING SYSTEM

Both games mentioned use a model that has been used for decades; a “mass spring system”. This is simply a collection of point masses connected by a series of springs, roughly in the shape of the object you desire. You can think of it like a particle system, where each particle is attached to some other particles by a number of springs. See figure 1 for a simple example.


Figure 1 – A simple mass-sping system with three masses and three springs

A simple spring connects two points, and has four basic parameters.

1. The rest length, which is the length of the spring when it is neither stretched not compressed
2. The minimum length of the spring when fully compressed
3. The maximum length of the spring when fully extended
4. The force exerted by the spring, proportional to its displacement from the rest length

Some springs can exert a different force depending on if they are compressed or stretched. The force can also vary in a non-linear relationship with the displacement, but for our purposes, the simple spring described above works well, and is easy to use.

A simple point mass has three parameters
1. The position in space, expressed as a 2D vector
2. Its previous position
2. Its mass

For most of what we are doing, I use a mass of 1.0 for all the point masses. However it’s useful to have a per-point mass, as it makes it easy to try various effects. If you end up with all the masses being the same, then you can obviously optimize this out of the computation.

VERLET MADE EASY

Verlet Integration is a fancy name for a slightly different way of applying the velocity and acceleration of a point to its position. Normally a point P will have a position X and a velocity V. Forces act on the particle, namely gravity, air resistance and the springs. The traditional (non-Verlet) way of updating the position of a particle is to first update the velocity with the acceleration and then update the position with this velocity:

F = total of forces acting on this point
T = Time step to update over
V += T*F/M
X += V*T + F/M/2*T*T

This generally referred to as Euler integration (with a second order Taylor series correction), but you might recognize it as the standard Newtonian equations of motion, (more usually notated as v=u+a*t and s=u*t+1/2*a*t*t). While referring to this as “integration” is technically correct, and will lead to a deeper understanding eventually, don’t worry if you don’t follow what is meant by “integration” – just think in terms of the equations of motion.

Verlet integration is basically another way of performing this physics advancement step. With Verlet we do not store the velocity, instead we store the previous position, and the velocity is implied as the difference of the current position from the previous position. The physics update then becomes:

F = total of forces acting on this point
T = Time step to update over
X0 is the previous position, X1 is the current position
XT = X1
X1 += (X1-X0) + F/M*T*T
X0 = XT

So why use Verlet, well technically using Verlet integration is more accurate than using Euler integration when the forces vary along with position and velocity. The reasons why this is so are a little obscure, and for many practically game purposes this difference in accuracy is not a major issue. The main reason for using Verlet is that it makes it very easy to apply constraints to a physical system. When a point moves past a physical limit (such as one point moving further away from another point than the maximum length of a spring that connects them), then we can simply move the point back to a “safe” position within that length. There is no need to calculate an impulse velocity, as the velocity is implied in the position, and is automatically handled by the movement.

BUILDING A BLOB

Once I got the basic spring-mass system working, I needed to create a blob. I figured that since the natural shape of a body of water is a sphere (when subjected to neutral external forces, such as a rain drop, or a blob of water in zero gravity), then I should start with a circular spring mass system, and then the application of gravity would naturally deform it into a nice blob shape.


Figure 2 – A single skinned blob, suffers from an easily folded skin

So my first attempt (figure 2) was a circle of 20 point masses, each joined to each other , and to a center point by springs. This is a standard N-gon, with the rest lengths of the springs being the natural lengths of the sides of the N-gon. This worked reasonably well for a first pass, and gave me something vaguely blobby that settled into circle under zero gravity, and deformed a bit when resting on the ground under gravity. But it suffered from several problems:
• The “skin” of the blob (the lines around the edge) kept folding over themselves, leading to ugly spikes in the skin.
• The blob was either too wobbly, meaning the edges of the blob wiggled like a giant piece of Jello, or too stiff, meaning it looked like a rubber ball.
• It kept getting caught on things, the outer edges would wrap around a corner of the environment, and the blob would hang there, looking like a dishrag.

My first attempt at solving this was to make the inner springs (the “spokes), have a longer rest length, so they would be under compression, and have the outer springs (the “skin”), have a shorter rest length. I was thinking this would simulate surface tension. However, this did not work very well, the shape did not improve, and the blob tended to violently collapse if gently nudged.

A BETTER BLOB

So I decided I needed a bit more structure to the blobs to make them more stable. After a few more failed experiments I hit upon the solution. Simply give the blob two layers of skin, one inside the other like concentric circles, joined together with a fairly rigid zig-zag set of joints. The inner skin is joined to a central point as before. See figure 3.


Figure 3 – Double skinned blob, the double skin structure provides a very stable skin.

This works remarkably well. I had to tweak the constants a bit, most specifically the number of segments, the thickness of the skin, and the strength of the springs. But quite quickly I had a very realistic acting blob. See figure 4 for the blobs in action

Why does this work so well? A “blob” here is a blob of very thick and slippery liquid, something like mercury. Mercury has a negative coefficient of surface tension, meaning the “skin” of a drop of mercury has very different properties to the interior. I initially though that the increased tension within the skin structure of our new blob was in some way simulating the effects of surface tension. But after looking at it for a while, I saw that the main effect it was having was constraining the curvature of the skin, thus smoothing out the high frequency wobbles we saw earlier. It’s simulating the appearance of surface tension rather than the underlying physics.


Figure 4 – The blob physics in action with various blobs. Download the sample to see for yourself.

BLOB PROBLEMS

I encountered three major problems in implementing this system. Firstly the blobs tended to be unstable, and wobbled all over the screen in the absence of external forces. Secondly, the blobs would get stuck, especially against corners, but also against surfaces. Finally the blob edges tended to get twisted when they impacted the environment at high speed.

The first problem (instability) struck me as a bit odd, since Verlet integration is known for being a bit more stable than Euler integration. This problem had me scratching my head for a while, but I finally figured out that it was due to the order in which I was performing the updates. I was looping over all the points, gathering the forces for a point (from the springs) and then moving the point.

The problem with that was that when a point moved, the force it applied to another point via a spring would change. In the case of two points connected by a single spring, the force exerted by the spring should be symmetrical for each point. However, if you move one point before gathering the force for the second point, then the forces will be different. This causes the spring system to have a net force imbalance in a particular direction (depending on the order of update). The solution here was very simple. I just split the loop up into two separate loops: one to gather the forces, and then one to apply them in the integration step. This ensured that all forces were symmetrical.

The second problem (getting stuck) has to do with the way collisions are handled. Since the collision of a point is implicit in its last movement, then if a collision resolution causes a point not to move very much, then it effectively kills its velocity. For collision resolution of a single point mass to work correctly (i.e. bounce off the surface), the next movement must be of appropriate magnitude and direction so future movement is correct. However, we are not simulating points; we are simulating a blob, so we need to consider the movement of the system as a whole.

With a spring mass system, the compression of the springs can handle the bouncing (to a certain degree). So if the leading edge points of a spring mass system simply stop when they hit a wall, the springs connecting to the points behind them will be compressed, and eventually bounce the whole blob off the wall in a nice convincing manner.

This works fine for something that just bounces up and down, but something hitting a surface at an angle needs to slide along the wall. This was quite easily accomplished with point/surface collisions by simply allowing the point the move parallel to the wall by the distance it would have originally traveled.

Something similar was done with line/surface collisions, but instead of the points moving parallel to the surface, they move parallel to the line. This allows the blob to slide over corners.

These collision resolutions were also where I implemented friction, simply scaling the distanced moved by a constant (like 0.95) gives a relatively pleasing result. You could calculate a friction force to be applied the next frame, but it’s simpler to directly incorporate it into the movement calculation. In the demo the friction can be altered by holding “S” to become slippery, and “A” to become less slippery. Holding “S” will allow you to slip though holes faster.

The final problem was edges getting twisted. This generally happened because a point moved furthur past another point it was supposed to keep away from. Since the spring constrain only measures distance, the point is then pushed away by the spring, but in the wrong direction, causing the edge to become twisted. One it’s twisted, it does not become un-twisted by itself.

The simplest solution, and the one I implement in the demo, is to try to never move in large steps. The easiest way of doing this is to run the physics multiple times with a smaller time-step. In the demo I run it six times per frame. Even so, the blobs can get kinks in them if shaken violently.

Something that exacerbates this problem is increasing the number of segments in a blob. With a large number of segments, the skin edges are much shorter, and so more likely to exceed their constraints in a single iteration. A lower number of segments works better. I found a 40 segment blob was impossible to kink, and yet still looked almost as nice as a 80 segment blob that was much more prone to kinking.

Obviously running the physics multiple times is not ideal, as it can be quite expensive. A better solution would be to simply ensure the kinking does not happen in the first place. Perhaps by adding some kind of angular constraint to a point on the surface. Another alternative is to link surface points to their second neighbors with a rigid constraint, so if the point gets past the first neighbor, then the second neighbor will push it back into the correct position. This type of second-neighbor constraint is commonly found in cloth simulation.

RESOURCES

Thomas Jakobsen, Advanced Character Physics, Gamasutra 2001, http://www.gamasutra.com/resource_guide/20030121/jacobson_01.shtml

Erleben, et al, Physics Based Animation, 2005, Charles River Media, Chapter 8, p265.

Chronic Logic, Gish Demo, http://www.chroniclogic.com/index.htm?gish.htm