Bendy Bones #49315

New Issue

Chris Offner · 2016-09-10T09:31:57+02:00

Chris Offner commented

2016-09-10 09:31:57 +02:00

System Information
Windows 10
NVIDIA GTX 980 Ti

Blender Version
Broken: blender-2.78.0-git.d0e7c7a-windows64

Bendy Bones flip strangely, head behaves differently than tail

Steps to reproduce:

Create Armature (Single Bone)
Subdivide bone into 3 bones
Move head of middle bone on its Y axis close to head of root bone
Move head of last bone on its Y axis close to tail of last bone
Rename root bone to Head, middle bone to Bone, last bone to Tail
Unparent Tail bone from Bone bone
Apply Bone Constraint StretchTo from Bone to Tail
In "Bone"'s bone settings, check "Use Custom Handle References"
Set In: to Head and Out: to Tail
Rotate Head and Tail separately on axes perpendicular to bone (i.e. bone's Z or X axes), compare behaviour of Head and Tail.

When following Jason Schleifer's tutorial to set bendy bones up with these very nice and intuitive controls, I get strange instability/flipping even before the head bone reaches 90 degrees rotation. The tail bone however behaves as expected.
I'd expect both of them to behave identically, i.e. like the more stable tail bone.

Demo 1 (Click to play)

Demo 2 (Click to play)

Demo 3
Moving Head off its rest position. Click to play.

Demo 4
Twisting Head and Tail respectively with Head off the rest position. Click to play.

Two separate project files showing the same setup (one horizontally in world space, one vertically in world space):
bendybones_vertical.blend

bendybones.blend

Charles Wardlaw regarding this issue asked:
Are you aligning the root and tip quats to a common vector before extracting twist?

**System Information** Windows 10 NVIDIA GTX 980 Ti **Blender Version** Broken: blender-2.78.0-git.d0e7c7a-windows64 **Bendy Bones flip strangely, head behaves differently than tail** **Steps to reproduce:** 1. Create Armature (Single Bone) 2. Subdivide bone into 3 bones 3. Move head of middle bone on its Y axis close to head of root bone 4. Move head of last bone on its Y axis close to tail of last bone 5. Rename root bone to *Head*, middle bone to *Bone*, last bone to *Tail* 6. Unparent *Tail* bone from *Bone* bone 7. Apply Bone Constraint *StretchTo* from *Bone* to *Tail* 8. In "Bone"'s bone settings, check "Use Custom Handle References" 9. Set *In:* to *Head* and *Out:* to *Tail* 10. Rotate *Head* and *Tail* separately on axes perpendicular to bone (i.e. bone's Z or X axes), compare behaviour of *Head* and *Tail*. When following Jason Schleifer's tutorial to set bendy bones up with these very nice and intuitive controls, I get strange instability/flipping even before the head bone reaches 90 degrees rotation. The tail bone however behaves as expected. I'd expect both of them to behave identically, i.e. like the more stable tail bone. **Demo 1** (Click to play) ![bendyBone.gif](https://archive.blender.org/developer/F356979/bendyBone.gif) **Demo 2** (Click to play) ![bendyBones2.gif](https://archive.blender.org/developer/F356986/bendyBones2.gif) **Demo 3** Moving *Head* off its rest position. Click to play. ![bendyBones3.gif](https://archive.blender.org/developer/F356992/bendyBones3.gif) **Demo 4** Twisting *Head* and *Tail* respectively with *Head* off the rest position. Click to play. ![bendybones4.gif](https://archive.blender.org/developer/F356998/bendybones4.gif) Two separate project files showing the same setup (one horizontally in world space, one vertically in world space): [bendybones_vertical.blend](https://archive.blender.org/developer/F356977/bendybones_vertical.blend) [bendybones.blend](https://archive.blender.org/developer/F356976/bendybones.blend) Charles Wardlaw regarding this issue asked: *Are you aligning the root and tip quats to a common vector before extracting twist?*

Joshua Leung was assigned by Chris Offner

2016-09-10 09:31:57 +02:00

Chris Offner commented

2016-09-10 09:31:57 +02:00

Changed status to: 'Open'

Chris Offner commented

2016-09-10 09:31:57 +02:00

Added subscriber: @chrisoffner3d

Chris Offner commented

2016-09-10 09:33:24 +02:00

Click the demo GIF images to show their full animation.
(The preview thumbnails are just static images it seems.)

Click the demo GIF images to show their full animation. (The preview thumbnails are just static images it seems.)

Jeffrey commented

2016-09-10 22:40:19 +02:00

Added subscriber: @italic

perfection cat commented

2016-09-11 11:37:38 +02:00

Added subscriber: @sindra1961

perfection cat commented

2016-09-11 11:37:38 +02:00

It did not occur if I excluded a check of Bone>Relations>Inherit Rotation of Bone.
Bone>Relations>Inherit Rotation: on

Bone>Relations>Inherit Rotation: off

It did not occur if I excluded a check of Bone>Relations>Inherit Rotation of Bone. Bone>Relations>Inherit Rotation: on ![bendybones_vertical_on.png](https://archive.blender.org/developer/F357702/bendybones_vertical_on.png) Bone>Relations>Inherit Rotation: off ![bendybones_vertical_off.png](https://archive.blender.org/developer/F357704/bendybones_vertical_off.png)

Joshua Leung commented

2016-09-20 16:16:05 +02:00

The "good" news here is that this isn't a regression caused by the Bendy Bones changes. The custom handles feature only makes a few changes to the inputs of the calculations performed here (but not the way that the calculations responsible for doing anything of interest happen). This can be verified by disabling the custom handle refs, or by loading the files in old releases. However, this means that we still have to fix whatever it is that's causing this issue...

AFAICT, what's currently happening is that the code works by:

Computing a "reference" rotation for the handle (using the handle position/offset, and "zero" roll). This uses the same code that's used for calculating the roll for bones when you create them - complete with all the quirks that can crop up there. You don't really want to mess around with this too much... what we've got now is a relatively good compromise between "mostly-working with predictable quirks" vs "random unpredictable failures". Reading up on that code does reveal however that there are some persistent numeric precision errors here that have cropped up several times in the past... (it's kindof a fickle "goldilocks zone" issue - too coarse and it gave errors, too fine and it gave errors)
Taking the final transform of the next/prev handle bone and transforming that into the current bone's space
Computing the delta transform between these two (i.e. how much the handle has "rotated")
Compute a roll value by using atan2f(diff.zaxis.x, diff.zaxis.z) (remember, bones use the "y" axis the direction of the bone, so the x and z axes define the roll, and since z is often used as the "up" axis, it's a good reference)

My first thought was that it might be due to some discontinuities in atan2f() (though that doesn't entirely make sense as those are at 180 degrees, instead of near 90 degrees). So, as a first step, I tried normalising the vector being used for this calculation. However, this just made the problem worse, by now causing a whole lot more flipping than was already happening (i.e. now at +/- 86 degrees, 139 degrees, and 202 degrees).

I then added some debug code to try and get a bit more insight into what was going on. It turns out that the roll values were going from about -0.085/-7.056 degrees (roll1/roll2) at z=85 degrees, to 1.70/-22.709 at z=86 degrees, to -142.99/-161.53 at z=87 degrees!

After mulling over it more, I realised that the Stretch To constraint here might also be contributing somewhat to the problems here. This is because the Stretch To constraint basically rebuilds a transform matrix from scratch, and may therefore end up not putting everything back together in the right way. This lead to the discovery of a workaround:

On the Stretch To constraint, change the Plane to "Z"

This workaround however is only really useful if you're only going to rotate on the Z-Axis and not the X-Axis (as it ends up flipping which axis exhibits these artifacts). So, probably not that useful in an actual rig for animators to use, but it at least gives us some options.

So, still more investigation needed to figure out what exactly is the real source of this problem, and what the best fix to apply would be...

The "good" news here is that this isn't a regression caused by the Bendy Bones changes. The custom handles feature only makes a few changes to the inputs of the calculations performed here (but not the way that the calculations responsible for doing anything of interest happen). This can be verified by disabling the custom handle refs, or by loading the files in old releases. However, this means that we still have to fix whatever it is that's causing this issue... ------- AFAICT, what's currently happening is that the code works by: 1) Computing a "reference" rotation for the handle (using the handle position/offset, and "zero" roll). This uses the same code that's used for calculating the roll for bones when you create them - complete with all the quirks that can crop up there. You don't really want to mess around with this too much... what we've got now is a relatively good compromise between "mostly-working with predictable quirks" vs "random unpredictable failures". Reading up on that code does reveal however that there are some persistent numeric precision errors here that have cropped up several times in the past... (it's kindof a fickle "goldilocks zone" issue - too coarse and it gave errors, too fine and it gave errors) 2) Taking the final transform of the next/prev handle bone and transforming that into the current bone's space 3) Computing the delta transform between these two (i.e. how much the handle has "rotated") 4) Compute a roll value by using `atan2f(diff.zaxis.x, diff.zaxis.z)` (remember, bones use the "y" axis the direction of the bone, so the x and z axes define the roll, and since z is often used as the "up" axis, it's a good reference) ------------ My first thought was that it might be due to some discontinuities in `atan2f()` (though that doesn't entirely make sense as those are at 180 degrees, instead of near 90 degrees). So, as a first step, I tried normalising the vector being used for this calculation. However, this just made the problem worse, by now causing a whole lot more flipping than was already happening (i.e. now at +/- 86 degrees, 139 degrees, and 202 degrees). I then added some debug code to try and get a bit more insight into what was going on. It turns out that the roll values were going from about -0.085/-7.056 degrees (roll1/roll2) at z=85 degrees, to 1.70/-22.709 at z=86 degrees, to -142.99/-161.53 at z=87 degrees! After mulling over it more, I realised that the Stretch To constraint here might also be contributing somewhat to the problems here. This is because the Stretch To constraint basically rebuilds a transform matrix from scratch, and may therefore end up not putting everything back together in the right way. This lead to the **discovery of a workaround**: ``` On the Stretch To constraint, change the Plane to "Z" ``` This workaround however is only really useful if you're only going to rotate on the Z-Axis and not the X-Axis (as it ends up flipping which axis exhibits these artifacts). So, probably not that useful in an actual rig for animators to use, but it at least gives us some options. So, still more investigation needed to figure out what exactly is the real source of this problem, and what the best fix to apply would be...

Demeter Dzadik commented

2019-08-03 01:12:55 +02:00

Added subscriber: @Mets

Joshua Leung was unassigned by Dalai Felinto

2019-12-23 16:37:00 +01:00

Dalai Felinto commented

2019-12-23 16:37:00 +01:00

Added subscriber: @JoshuaLeung

Sybren A. Stüvel commented

2020-01-30 14:52:58 +01:00

Added subscriber: @dr.sybren

Sybren A. Stüvel commented

2020-01-30 14:52:58 +01:00

Changed status from 'Confirmed' to: 'Resolved'

Sybren A. Stüvel closed this issue

2020-01-30 14:52:58 +01:00

Sybren A. Stüvel self-assigned this 2020-01-30 14:52:58 +01:00

Sybren A. Stüvel commented

2020-01-30 14:52:58 +01:00

Unfortunately the Vimeo link is down, so I removed it from the description. The uploaded GIFs are still good.

I can reproduce this issue in 2.79, but can't reproduce this issue any more with Blender 2.82 alpha (87b551e836) so it appears the bug has been fixed already.

Please try the latest daily build: https://builder.blender.org/download/

If the problem persists, please let us know so we can re-open the report. Don't forget to mention the specific version you tested again.

Unfortunately the Vimeo link is down, so I removed it from the description. The uploaded GIFs are still good. I can reproduce this issue in 2.79, but can't reproduce this issue any more with Blender 2.82 alpha (87b551e8365954d03d1c27303b9776deefd4e4d3) so it appears the bug has been fixed already. Please try the latest daily build: https://builder.blender.org/download/ If the problem persists, please let us know so we can re-open the report. Don't forget to mention the specific version you tested again.

Sign in to join this conversation.

No Label

Download

What's New

Blender Studio

Manual

Developers Blog

Documentation

Benchmark

Blender Conference

Development Fund

One-time Donations

Bendy Bones #49315