This is an old revision of the document!
Animations are a central point in any game as they give life to 3D models. When it comes down to animation systems there exists in general three types: animation sequence, parametric animation and procedural animation. Furthermore animations can be classified as static or dynamic. The animator system combines all three types of animations providing a rich set of possibilities to create all kinds of animations. The following list describes briefly the three animation types.
The animation sequence is used often for titles where complex animations are not required. There an animation sequence created by an animator is simply played back. This is the most simple form of animation requiring the least processing power. Depending on the game type animation sequences can work well. The Drag[en]gine provides support for animation sequences using the Animation resource type. Animation resources store a list of moves which contain each one specific animation sequence. Moves store the animation sequence for a set of bones. These bones are matched on run-time against the bones of an actual 3d model rig whereas non-matching ones are ignored. Animation sequences are therefore a static animation type.
The parametric animation is a step up from the simple animation sequence in that multiple animation sequences are blended together to create a dynamic animation. The blend weights determine how much influence one animation sequence has on the final result. This animation type is the de-facto standard for triple A game engines looking for complex and highly dynamic animations. Usually these systems are accompanied by a pre-processing engine which determines transition points between the individual animation sequences. These systems can become quickly very complicate to work with since you need all kinds of animation sequences for the possible variations on the animation a player can provoke. Usually this system is used together with motion capturing since creating all the animations from hand can require some time (often needed though… you can't mo-cap a dragon for example). Parametric animations provide in general good animation results especially with large animation sets. This is though also the problem that you need many animation sequence variations to build believable dynamic animations. You can not simply create a new pose without modifying large parts of the animation set.
Procedural animations are the most dynamic type of animations often not requiring animation sequences at all. Here movie CG animation techniques like inverse kinematic or animation constraint systems are used to calculate the pose of a character on run time. The absence of any limitations as they exist with parametric animation type makes procedural animations powerful but also limited at the same time. Procedural animations have a tendency to be robotic and not life like since they are driven by mathematics instead of an artistic hand (or mo-capped). Often procedural animations are used only to adjust parametric animations to give the illusion of a larger set of animations. Pure procedural animation systems are usually not often found due to the mentioned limitations and problems with this technique (depends though a lot on the game in question). Physics based animation also belongs into this category since the principle is the same just that physical functions are used.
The main idea behind the animator system is to separate the animation logic from the animation production. In most systems these two parts are intervened. While this has advantages it also imposes limitations and locks game developers to a given type of animation. The animator system in contrary provides the game designer with a simple to use system where he decides how the animations are build instead of using a pre-build system. To use an animator you have to assign it a component upon which the produced animations are applied. Optionally you can assign the animator an animation which will be used by some rules to retrieves animation sequences from. This is not required if you do not plan to use animation sequences but only procedural ones. Once setup you can apply the animation to the assigned component. Using the source and destination blend factor you can blend the animation state produced by the animator. This way you can either blend a calculated animation with a state calculate by the physics system ( for example make an unconscious body regain consciousness ) or apply multiple animators in succession. This can be used to separate different kinds of animation scenarios into manageable units ( one animator per scenario ) doing transitions between them. You can also skip animators altogether and apply your self calculated bone states to the component directly. Using animators is though the way to go since this way advantage can be taken of future hardware solutions as well as better animation algorithms.
Controllers provide the connection between the animation logic and the animation production. The game developer uses in his game a given set of variables to track the state of an actor for example. The idea is to allow the game developer to map these variables directly to the animation system not having to worry about converting them into whatever system one might use. This is done using the controllers which act as dial knobs the game developer can mock around with like on a mixer. Each controller has a minimum and maximum value. The current value of the controller is always kept inside this boundary. The game developer can set here the limits of the values he uses in his variables. He is not required to normalize or otherwise apply fancy math to the values in his game to match them with the animation system. It's the animator system that adapts to the game developer. Controllers can either clamp or wrap values outside the limits. Clamping is typically used for controllers attached to actions like raising or lowering the head where you want the motion to stop at the extreme points. Wrapping on the other hand is used if the controller is incremented with respect to the elapsed time. Hence using clamping an animation would stop at the end while with wrapping it would loop forever. The game developer is responsible to increment the value of a controller if it is supposed to be driven by the elapsed time. This can be done easily using an increment function which does all the clamping or wrapping for you. In addition controllers can be frozen which prevents their current value from changing. This way the game designer is not required to keep track in his game when to manipulate a controller and when to spare it. On a frozen controller all methods intended to change the value will be silently ignored. It is fully valid to call such methods on a frozen controller. The controller normalizes the current value itself. A particularity about the controllers is that they own in addition to the single value also a 4-component vector value. This value is used by specific rules and is neither clamped nor altered. The order of the controllers is important since they are referenced by their index later on.
Controllers deliver a linear and normalized value. This is though not always the desired result. For example an animation sequence could be desire to have full influence for a controller value or 0 decreasing linearly down to no influence at the point 0.5 while staying at no influence up to the value 1. Links take on the duty to map controller values to actual weight values which can be consumed by rules later on. A link consist of a piecewise linear curve which maps incoming controller values ( x coordinate ) to weight values ( y coordinate ). Furthermore the source controller is indicated using the index of the controller ( for speed and handling reasons not a pointer ). A link is valid if the source controller is a valid index of a controller otherwise it is an empty link. Vector values of controllers are not altered by the mapping curve in any ways and are handed to the target as they are. The order of links is important since they are referenced by their index later on.
Targets define the sink of values delivered by links. Every rule has one or more targets referring to attributes that can be dynamically altered. As with controllers the links are referenced by index instead of a pointer for the same reasons. A target can have any number of links assigned. If more than one link is assigned the value delivered by them are multiplied to produce the final weight value used by the attribute in question. Only valid links influence a target. A target is bound if it contains at least one valid link otherwise it is unbound. For a bound target the attribute in question takes on the value of the target. If the target is unbound a default value ( usually stored separately in the rule itself ) is used. This means especially that if you have links assigned to a target it is still possible that the default value is used if none of the links is valid. Targets expecting a vector take the vector of the first valid link. No merging of vectors of multiple links is conducted.
Rules provide the production rules for an animation. There exist various rules each belonging to one of the three animation types. All rules together have common properties. Each rule can be enabled and disabled individually which allows to modify the animation production process without altering the rules itself ( which is faster ). The result of each rule is blended with the existing animation state using the source and destination blend factor. The rule animation state is multiplied with the source factor and added to the existing animation state multiplied by the destination factor. For both the blend factors there exists an appropriate target. This can be used to alter the influence strength of a rule which is useful to gradually blend in animations ( for example an idle animation or typically used on inverse kinematic rules to avoid jumping ). By default the source factor is set to 1 and the destination factor to 0. A typical setting is 1 for both in which case the animation state of the rule is added to the existing one ( used for doing parametric animation setups ). In addition each rule has a list of bones. This list indicates which bones are affected by this rule and can be used to apply rules only to subsets of a character ( for example torso and legs separately ). If the bone list is empty this equals to the rule influencing all bones. The order of rules is important since they are applied in the order they are specified. The following list contains all existing rules.
The animation rule is a static rule and applies a frame taken from an animation sequence. You can specify the name of the move to take the frame from as well as the time of the frame. If the move can not be found the default state with zero position, orientation and size for all bones is applied. The time is clamped to the playtime of the move. For the time attribute a target exists. The value of the target ( in the range from 0 to 1 ) is mapped to the range from 0 to the playtime of the move. The same clamping applies for the target.
The animation difference rule works similar to the animation rule in that it takes frames from an animation sequence. The difference is that two frames are queried at the same time and the difference between both becomes the result of the rule. This can be used to apply an animation relative to the existing one ( for example a turret motion ). For both frames the move name and the time can be specified. The same rules apply as for the animation rule. Care has to be taken to not obtain zero quaternions which happens if both frames have the same orientation. Usually this rule is used with the source and destination blending factors 1 and 1 to add the difference to the existing animation. In this case zero quaternions are no problem. For both the leading ( the frame to take the difference from ) and the reference ( the frame to subtract from the former ) frame time there exists a target which behaves as the one in the animation rule.
The bone rotator rule is a procedural rule which allows to rotate bones relative to a coordinate frame. This is useful to apply motions like turning the head or torso left and right ( or up and down works too depending on the character ). A minimum and maximum orientation in Euler angles has to be provided. The rotation target value determines the orientation to be picked with 0 retrieving the minimum orientation and 1 the maximum orientation ( with all other values a linear blend between the two ). By default the rotation takes place around the center of each bone ( each bone is operated on individually and independent of all others ) with the orientation of the coordinate frame equal to the one of the component. This simply rotates the bones around their position relative to the component. If bone local rotation is disabled the bones are rotated around the component center instead. Since the rotation is specified in non-clamped Euler angles you can specify rotations outside the 90 degree range without obtaining strange results.
The inverse kinematic is a well known tool for movie computer graphics and is the opposite of forward kinematics ( which is named here animation sequences ). In forward kinematics for each bone a state is given and according to this the rig is posed. The problem with this technique is that ( although fast to compute ) objects do not necessarily touch other objects. A typical symptom is the 'sliding feet' ( or also called 'shuffling feet' ) problem where the feet of a character slides on the ground instead of sitting tight to it. Inverse kinematics presents a solution to this problem as you provide a point ( and optionally orientation ) in space a given bone has to reach. The animator module then tries to calculate for the bone in question and all bones up to a base bone the rotations required to achieve the given goal. The inverse kinematics technique sports its own problems as bending into the wrong directions or various forms of 'gimbal locks' ( goals that are not reachable or where way different poses achieve the same goal which results in snapping motion ). Especially with larger IK chains wrong poses can happen quickly. Another problem is the way of calculating the bone positions. For doing inverse kinematics usually an iterative process is used. During each step the bones are slightly nudged to get the target closer to the goal. If the error is less than a threshold the process stops and the IK chain is considered resolved. This still induces a sliding feet but it is too small to notice. Due to the iterative nature of the process though calculating inverse kinematics rules is usually not cheap. For an inverse kinematic rule to be valid two bones are required to be specified. One is the target bone which has to reach the goal and the other is the base bone which is the last bone in the chain between the target bone and the base bone which is altered. Both bones have to be either parent and child or connected by a series of bones. Furthermore the position and orientation of the goal can be specified as well as the position and orientation of the target point. The target point is relative to the target bone and allows to position a bone relative to a specific point ( for example the finger tip ). There is only two targets which are the goal position and orientation. Both targets expect a vector and therefore only the first valid link delivers the vector. The goal position and orientation is relative to the component. Therefore if you want to plant feet on the ground you have to transform the contact point in world coordinates first into component local coordinates.
The state manipulator rule alters the state of bones directly. They are very fast since they neither query an animation frame nor do they apply complex calculations. For each the position, orientation and scaling a minimum and maximum value can be specified. As with the bone rotator the individual attributes are blended between the minimum and the maximum using the target value ( one target for position, orientation and scaling ). This is the typical rule for animating eyes or turrets where only one bone is altered ( typically the orientation only ). The manipulation of each attribute can be enabled and disabled separately.
The state snapshot rule stores a static animation state which has been obtained at some time in the past. The state can be initialized either from an animation frame or from the current animation state of the component. This usually used to blend gradually from a position obtained from a physics simulation to an animation state ( for example getting up from fall on the ground ). At the start of the transition process the current state would be snapshotted. From there on using the blending factor targets the state can be transformed gradually into the other animations. This can also be used to freeze a produced animation in time ( for example encasing an enemy in ice ). No additional targets are defined by this rule
The foreign state rule applies the state of a specific bone to the influenced bones. This can be used to transfer a state from one bone to another. Typically these are bones which are not animated otherwise as for example attachments on clothing which are applied to an actor and should animate reasonably with it. You can determine relative to which coordinate system ( bone or component ) the state is obtained and and applied. It is possible to mix those ( for example obtaining the state relative to the component but applying it relative the the bone local coordinate system for the target bones ) but usually for both component space is used. You can specify for each attribute individually a scaling factor with which the incoming bone state is first multiplied. This scaling factor also owns a target for each attribute. Scaling can be used to weight multiple bones into a new one. For example if you have an attachment which is located halfway between two animated bones you can use two foreign state rules one for each bone scaling them by 0.5 and adding the result to obtain the final bone position. Each attribute can be enabled separately which prevents altering the appropriate attribute.