mocap for non-humanoid rigs — quadrupeds, creatures, anything weird

the proxy skeleton idea is solid but the hidden cost is you now have a third rig to maintain. any time the quadruped rig gets updated you're potentially breaking two retargeting layers instead of one. did this on a dog character last year and it was fine until our rigger added a spine joint and everything downstream needed rekinking. worth it on a bigger project, just go in knowing the maintenance overhead, especially if the rig is still evolving when you start capturing.

tbh this is the thing people skip straight past when they say "just keyframe it." you can get limb positions close enough by hand but the weight-shift timing from an actual performance — even someone crawling around awkwardly on camera — has a quality that's genuinely hard to manufacture from scratch. it reads as alive vs animated. not saying keyframe is bad, just that this is the specific thing you're actually trading away.

The center-of-mass point is real and I think it's underrated in these conversations. The spine and hip curves from a human performer, even crawling around awkwardly, carry timing information that maps surprisingly well to medium-sized quadrupeds. Weight shift on footfall, hip counter-rotation, recovery timing, all of it tends to read credibly once it's on the right rig.

The thing that tends to break it is lateral sway. Human balance compensation produces a side-to-side hip motion that's pretty specific to bipedal locomotion and just reads wrong on four legs. If you identify and correct that curve specifically before you hand-animate the limbs on top, you save a lot of iteration time downstream. It's a narrower fix than re-doing the whole hip animation.

yeah timing is exactly what kills it. one workaround i've seen is capturing on a treadmill so the actor is actually moving at pace, which forces a more natural rhythm even if the limb shape is still off. doesn't solve the skeleton mapping problem but at least you're not manually stretching every curve afterward to hit the right stride tempo. still a hack but a less painful one than trying to retime everything in the graph editor after the fact.

yeah this burned us. tried the proxy approach on a creature rig and the modeler added two extra spine bones in week 3, which meant redoing both retarget setups. if you're going the proxy route, lock that skeleton definition down before the final rig is signed off. otherwise you're paying the maintenance cost twice and the proxy stops feeling like a net win pretty fast.

the timing mismatch is the real killer though. human bear crawl is way slower and more deliberate than an actual dog trot, so even when the pose shape is usable the rhythm feels completely wrong. i ended up doing a manual time-warp of the mocap curves in blender against a reference video of the actual gait, tedious as hell, but the result felt way more alive than just scaling playback speed. if anyone's found a less manual approach to matching mocap timing to reference footage i'd genuinely love to hear it.

This split makes a lot of anatomical sense. Human center-of-mass dynamics actually transfer pretty well to quadrupeds. Weight shift timing, hip-shoulder counter-rotation, recovery from footfalls all read correctly when captured from a human performer. The limbs are where anatomy betrays you, since joint placement and segment ratios are so different. Keeping mocap for the trunk and hand-keying limb placement plays to both strengths.

The tricky part is making sure the hand-key pass is driven by the mocap root, otherwise your animated paws won't respond to the weight shifts happening underneath them. Constraining limb IK targets to proxy objects that follow the mocap trajectory keeps it feeling like one performance instead of two animations glued together.