[In this Gamasutra technical article, originally published in sister publication Game Developer magazine, programming veteran Brown shows how the Nintendo DS can be used for surprisingly complex character-based physics calculations using Verlet intergration.]
While the Nintendo DS enables some of the most creative and unique gameplay currently available, it doesn't exactly have a reputation as a computational powerhouse.
Developing a game for such a platform represents a unique challenge -- having a restricted budget of resources is not new to game programmers, but trying to innovate on a tight platform that you hope can coexist with the gaming experiences provided by the (comparatively) unlimited resources of the next-gen consoles requires a bit of bravery.
It is very true that, for average gamers, creativity rather than technological mastery forms the basis of quality for a game on the Nintendo DS. However, it is a very satisfying experience to successfully push the boundaries, not only of the technology, but also of the expectations of performance on a limited device.
For me, it was my own restricted expectation of performance, rather than the lack of computational power, that presented the largest barrier to implementing ragdoll physics on the Nintendo DS. The little brother of gaming actually held its own very well. The main idea that I used for the ragdoll was the common Verlet integration technique.
Ragdoll is a procedural animation technique that represents the interplay of physics and animation. On the physics side, you will need to have decent collision detection. Collision does not need to be too complex. My collision consisted of static sphere vs. sphere intersection tests for collision with objects, and a dynamic swept sphere vs. triangle test for collision with the world.
The triangles that composed the collision mesh of the world were organized by spatial partitioning with a K-D tree. I did not concern myself with self-intersection of the ragdoll, and it turns out that I didn't need to.
In terms of the animation system, you will need to be able to use the information you get from updating the physics on the ragdoll to specify the position and orientation of different joints on the body. If you are using a proprietary animation system, you will need to determine the best place to insert this information.
Generally, the transformations of the joints are represented in local space-meaning the transformation is relative to the parent joint-until all key frame and animation blending calculations have been made. Before skinning an animation, the transformation of the joints must be converted to animation space-relative to some character origin-or to world space.
Generally the code block where this conversion takes place is the best spot to inject the ragdoll information. If you don't have an animation system, or you feel like modifying your proprietary one will be too hard, you can use the animation system that comes with the Nitro SDK. This animation system has an appropriate callback that allows you to specify the position and orientation of a joint in animation space.
When the state of a particle is represented by a position and a velocity, the next position and velocity can be determined given a particular acceleration. The deviation in the trajectory has been exaggerated in order to show how integration introduces error.
The basics of a Verlet particle system are documented very well in Thomas Jakobsen's paper "Advanced Character Physics" (which you can find at www.teknikus.dk/tj/gdc2001.htm). The Verlet particle system is ideal for representing a ragdoll skeleton because constraints can be infinitely rigid without causing the system to diverge, thus the distance between joints can more rigidly be maintained. For this reason, you may find that a Verlet particle system is a nice thing to have around, even if you are not planning on doing any ragdoll physics.
The main idea behind the Verlet integration approach to ragdoll is that the positions of the joints are determined by point particles. The "bones" of the skeletons, represented by length constraints that enforce two neighboring particles, are an exact distance apart.
Extra length constraints can be added as necessary to prevent unnatural poses, as well as self intersection of the ragdoll. Updating the positions of the particles is relatively inexpensive, and calculating a length constraint is the equivalent of resolving a sphere collision, so the overall cost of updating the particle system is pretty low.
Once you've updated the positions of all of the particles, and then applied the constraints to ensure the appropriate distance relationships between them, it is time to do some collision. I decided to allow the particles to collide with the world as independent spheres, rather than trying to construct a closed collision volume that encompassed a ragdoll limb.
This decision was based on two factors. The first was the cheapness factor-sphere collision is much cheaper than some kind of ellipsoid or oriented bounding box collision. The second was that I succeeded in convincing the artists and designers who were creating the collision worlds to "play nice" with the ragdol -- meaning that they should try to make the static collision geometry as smooth as possible, with no really sharp kinks or protrusions that the ragdoll might get hung up on.
However, I still ran into pretty big problems using this approach. If you are using spheres to represent the collision volumes of joints, then the spheres generally need to be pretty small. After all, they are representing things like a hand, shoulder, or hip. Using such small spheres, I immediately began to notice that all of my collision routines would generally fail, due to the numerical imprecision introduced by fixed-point arithmetic. I'll cover some of the things that I learned in order to overcome these precision-related issues.