Scoring weights can now be specified at several levels of abstraction.
Channel: Pose weight vs trajectory weight.
Horizon: History weight vs prediction weight.
Type: Weights per feature types, such as positions and velocities.
Weights are normalized at each of these levels. So ultimately the weights that
make up all channels of a weights group will sum to 1 (or zero if entirely disabled).
Several new concepts were introduced for this system that aren't fully
formalized. Channels aren't a concrete data type, but I'm imagining we'll head
in that direction where a Schema is made up of Channels which themselves contain
sampling paramters like time/distance offsets and what to sample. The weights
system sets itself up in such a way to effectively pretend channels exist.
Similarly, horizons aren't an explicit data type either and the weights system
pretends they sort of are for presenting this high level concept of past vs future
weight to the user.
I have removed the previous weighting system including per sequence weights.
This feature will be replaced by a grouping system for databases in the future.
Groups will be collections of sequences and a set of weights for that group.
This is why the WeightsContext contains an array of 1 group of weights-- it's
setup that way to vector the implementation toward groups.
Speaking of the WeightsContext, the motion matching node holds onto one and
updates it every frame. This is in order to support changing weights at
runtime. There are a set of dynamic weight parameters that can be used to scale
the weights of all weight groups within a database. The WeightsContext tracks
the prior dynamic weight params and knows if the weights buffer needs to be
recomputed or not. So most of the time, updating the WeightsContext is a no-op.
In the future we'll need to add support for interpolating changes to dynamic
weights to avoid abrupt scoring changes.
#preflight 614452dd568b3a00013ee9cb
#rb cesar.castro
#jira none
[CL 17555406 by braeden shosa in ue5-main branch]
Included features in this changelist:
* Capture a snapshot of common movement properties (velocity, speed, etc) that are used to drive animation
* Character movement component for stop/pivot location
* Library of anim node functions for Distance Matching
* Template anim node that can advance by distance traveled rather than by time
* Template anim node that's driven by distance to a target rather than by time
* Animation modifier that generates distance curves based on root motion
* Turn in place functionality for keeping the capsule from spinning the pose when it rotates by applying an offset, includes functionality for turn on spot animations to compenstate for the offset
#preflight 61446e46599fd80001c250e2
#review-17532442 @Koray.Hagen, @Thomas.Sarkanen, @Fernando.Coello
#fyi Laurent.Delayen
[CL 17551436 by Aaron Cox in ue5-main branch]
#rb braeden.shosa, aaron.cox, cesar.castro
#jira none
Motion Trajectory Component for Motion Matching:
Motion Trajectory Component notes:
1) Abstract component/interface implemented with prediction and history API.
2) Implemented uniform, frame-rate independent history sampling algorithm for retaining trajectory sample coherence.
3) Implemented concrete Character Movement Trajectory Component for encapsulating ground locomotion prediction algorithm and API.
4) Motion Trajectory blueprint library containing:
5) FlattenTrajectory2D algorithm for isolating and removing Z axis direction contribution from tracjectory.
6) ClampTrajectoryDirection for projecting trajectory samples into a discrete, allowed set of directions (such as cardinal).
Engine changes:
1) Implemented shared TrajectorySample, Range, and Domain structures for common usage among Motion Trajectory, Pose Search, and Blueprint.
2) Expanded Sequence Player blueprint library functionality for Scripted Motion Matching content example.
Pose Search changes:
1) Match Pose renamed to Pose History with minor configuration parameters.
2) Complete removal of approximated past trajectory sampling. This has now been pushed to the Motion Trajectory Component.
3) Motion Matching interface changes: Goal has been replaced with Trajectory and Dynamic Play Rate Adjustment has been encapsulated to an experimental code module and blueprint library. New defaults, and encapsulating settings.
4) Motion Matching changes: New Pose Search blueprint library which encapsulates core Motion Matching algorithm.
5) Unified Trajectory to Pose Feature Vector transformation algorithms.
6) Removal of temporary Pose Search Prediction blueprint library (now Dynamic Play Rate).
7) End-to-end implementation support for Past/Future Time and Distance domain schemas.
AnimInput changes:
1) Removal of AnimInput_CapsulePrediction (replaced with Motion Trajectory Component).
Gameplay Anim Gym changes:
1) Pose search database is now fully distance-based (history and future).
2) Character blueprint includes Character Movement Trajectory Component with distance-based domain setup.
3) Scripted Motion Matching anim node function prototype with Engine content example.
#preflight 6143ef4681695600011c49a8
[CL 17547347 by koray hagen in ue5-main branch]
Also fixes type conversions (first time I have tried to convert to a base class and found that my IsChildOf check was the wrong way around).
Plus some extra accessors added for exec/update contexts.
Found the need for these when doing show & tell prep.
#rb Jurre.deBaare
[CL 17367354 by Thomas Sarkanen in ue5-main branch]
Fixes anim task blocking on RB sim task and doing 0 work, now taskgraph schedules work and doesn't lock up extra threads.
#jira UE-120634
#rb chris.caulfield
[CL 17083918 by Max Whitehead in ue5-main branch]
Crash using ControlRig in Sequencer
#info save/use cached pose nodes are not correctly invalidated/refreshed due to ControlRig edit mode invoking evaluation of the same anim instance multiple times per frame
#fix ensure that various counters used for synchronization/invalidation of cached data are used by ControlRigLayer proxy - and applied whenever update/evaluating the source "input" proxy
#ROBOMERGE-SOURCE: CL 16927983
#ROBOMERGE-BOT: (v838-16927207)
[CL 16928005 by jurre debaare in ue5-main branch]
#info save/use cached pose nodes are not correctly invalidated/refreshed due to ControlRig edit mode invoking evaluation of the same anim instance multiple times per frame
#fix ensure that various counters used for synchronization/invalidation of cached data are used by ControlRigLayer proxy - and applied whenever update/evaluating the source "input" proxy
#rb Martin.Wilson
[FYI] Mike.Zyracki, Halfdan.Ingvarsson
#ROBOMERGE-SOURCE: CL 16922496
#ROBOMERGE-BOT: (v836-16769935)
[CL 16926362 by jurre debaare in ue5-main branch]
For now, removing these as their semantics are due to change
Also converting Update and Become Relevant functions to use more-devived-typed update context (less casting required)
#rb Jurre.deBaare
[CL 16771205 by Thomas Sarkanen in ue5-main branch]
Fixes a case in RBAN where mesh component initialisation may pass through a zero delta time which can be used to divide in the calculation of a velocity when the node is set to inherit bone velocities.
#rb Benn.Gallagher
#jira UE-118618
[CL 16755554 by RyanyiNF in ue5-main branch]
- SetPreviewSkeletalMesh to UAnimationAsset and UAnimBlueprint
- Functionality to retrieve UAnimationGraphs and UAnimationGraphNodes of a specific class from UAnimBlueprint and UAnimationGraph respectively
- Functionality for pose driver/asset scripting
#jira UE-117797
#rb Thomas.Sarkanen
#preflight 60d075e32ab218000115936a
#preflight 60d1a3eaa819040001397a52
[CL 16741838 by Jurre deBaare in ue5-main branch]
Split 'node context' into an 'evaluation context' and a 'node reference' type and updated regular function signatures when creating node functions. This way we can seperate concerns (and potentially reference anim nodes standalone in the future).
Moved function caller code into function ref file to stop AnimNodeBase bloat.
Fixed functions not being called on root nodes.
#rb Jurre.deBaare
[CL 16727107 by Thomas Sarkanen in ue5-main branch]
Adds 'template' anim BP concept. These anim BPs have no TargetSkeleton and as such cannot have direct references to animations placed inside of their anim graphs
Adds function call support from anim nodes. All anim graph nodes can now call functions when initialized, updated, evaluated or when they first become relevant.
Relevancy is established using a new FAnimSubsystemInstance_NodeRelevancy which tracks nodes in a local map, per UAnimInstance, if nodes require it.
Functions are displayed on the node if bound, and in the details panel.
Added a new override point to FAnimSubsystemInstance to allow for WT init.
Moved FMemberReference customization into a public header so it can be used on anim node functions.
Wrapped functions up into FAnimNodeFunctionRef structure so they can be re-used more effectively. Converted CallFunction node to use them.
Added a couple of simple BP function libraries to demonstrate the use of the new FAnimNodeContext (this is intended to be a generic way of exposing FAnimNode_Base types to script without the node types themselves having to be known to script directly).
Added the ability to set exposed properties as 'always dynamic' so they appear in mutable data rather than being merged into constants. This allows for the anim node data API to be changed to be less strict (and more script friendly). Now values can be set in mutable data (via GET_MUTABLE_ANIM_NODE_DATA_PTR) and attempted sets to constant data can be ignored and reported to client code.
A few minor crash fixes with edge cases (inc when trying to open an asset editor fails because of a missing skeleton/cancellation).
Also fixes an issue where literal linked anim graph/control rig/custom poroperty node inputs didnt get copied
#rb Jurre.deBaare,Aaron.Cox
[CL 16703644 by Thomas Sarkanen in ue5-main branch]
