Retarget a Supplied Animation to Your Pitch Character Without Losing Contact
Retarget a Supplied Animation to Your Pitch Character Without Losing Contact
A retargeted clip almost always plays. That is what makes the failure easy to miss. The motion lands on your character, the timing survives, and at thumbnail size the pose looks like the pose. Then you scrub to the hold and the palm is hanging a hand's width from the switch you built into the set, or the support foot has been drifting backwards for eight frames while your character leans into a contact that is not there.
The contact is the thing you are actually transferring. So the workable sequence is: name the contact and the events around it, confirm that the two rigs correspond where it matters, run the transfer uncorrected and keep that version on file, then compare two repairs — limited correction, and a purpose-built action — under rules that keep the comparison honest. Judge the exported clip rather than the preview, and make the call with the export in front of you.
One question sits before all of this: the motion sample has to be one you are entitled to use. That is not an animation problem, and nothing below resolves it.
The contact is what you're transferring
In a reach-to-a-panel action, the contact is not the whole animation. It is a small, nameable set: the object, which is fixed in the set and does not move to suit a new character's proportions; the plant, which is a specific foot at a specific point on the floor; the arrival, which is the frame the palm meets the surface; the hold; and the release. Around those, the relationships that make the action read: where the weight sits, what the torso is doing, where the character is looking.
Watch the whole clip through once, at speed, before you examine a still. The arrival frame will look convincing in a screenshot. The slide that ruins the shot happens in the six or eight frames leading into it, and a still cannot show you a foot travelling. Mark your event frames on the source clip before you transfer anything. You are going to need a shared timing reference, and reconstructing it later from memory is how comparisons go soft.
Check correspondence before you tune anything
Epic's retargeting overview and its biped walkthrough describe the process as mapping chains — whole joint chains, matched between different skeletal structures — rather than matching two skeletons bone for bone. The walkthrough separates skeleton setup, chain correspondence and pose correction, and names the pelvis or hip as a typical retarget root for a biped. That structure is the part worth taking seriously. The step-by-step menus are not, because the overview page I'd be pointing you at showed Unreal Engine 5.8 and vocabulary that does not match the walkthrough's Chain Mapping and Retarget Root labels. Reconcile the two against the build you have open before you follow any instruction literally.
Then check the things that similar names hide. Where does the target's arm chain start and end — does it include the clavicle on one rig and not the other? Does the leg chain reach the toe or stop at the ankle? Where is the retarget root, how is it oriented, and what does the baseline pose look like on both characters? Does a unit scale differ between the two (a rig authored in centimetres in a metre scene produces a character who is a hundred times too large, or a transfer that looks almost right at exactly the wrong size)?
This is also where you decide whether the transfer can work at all. Two rigs posed in the same joint rotations trace different arcs if their bones are different lengths. That is not a bug in the transfer; it is arithmetic, and the rest of this article is about the arithmetic.
Run the raw transfer and keep the failure
Before you correct anything, export the uncorrected result and record what it does. Which direction does the support foot drift, how far, over which frames? Where does the hand stop relative to the contact point — short, long, low, high? What is the wrist doing, and does the palm meet the surface at the angle the contact implies? How far has the pelvis moved from where the source had it?
Then sort the observations into two piles. A constant offset — the whole clip rotated, the character consistently a few centimetres above the floor, every limb pointing the same wrong way — is a correspondence problem, and correcting contact will not fix it. A drift that accumulates only during the reach, or a miss that appears only at the contact, is a contact or performance problem. The two need different work, and the raw clip is the only honest baseline you will ever have for either. Save it as its own asset. If you correct first and keep nothing else, you will never be able to tell whether the corrected version is good or merely better than a version you can no longer see.
Fix the comparison rules before you judge any route
Before you build the second variant, write down what stays constant. The contact point stays where the set puts it. The arrival frame stays put. The review camera stays put, and you review the same intervals on every route: approach, hold, release. If the camera changes between variants, you have started grading two different shots.
What each route is allowed to change is the part people fudge. The raw route changes nothing. The corrected route may add goals or solvers and may change joint rotations, but it must keep the same plant and the same foot contacts — if you move the foot to make the hand reach, you have replaced the action rather than corrected it. The reanimated route may rebuild the approach entirely, because that is what reanimation means, but it must still put the palm on the same point at the same frame. Without that rule the reanimated variant wins by moving the problem somewhere the camera cannot see it.
A constructed comparison, with the arithmetic in the open
The rigs below are invented for this exercise — no engine, no transfer and no export was run to produce them — but the geometry is fixed and checkable, which is the point. You should be able to say why an action fails, not only that it did.
The source, call her Mara, is the character the supplied motion was authored for. The pitch character, Pell, is shorter and differently built. Both reach the same contact point on the same fixed panel.
- Contact point: 1.55 m above the floor, 0.48 m forward of the reaching shoulder in both poses.
- Reaching shoulder: Mara's sits 1.42 m up; Pell's sits 1.34 m up.
- Shoulder to palm, along the arm chain: Mara 0.71 m; Pell 0.58 m.
Mara's contact sits 0.13 m above her shoulder and 0.48 m in front of it, so she needs 0.50 m of arm — about 70% of what she has, a relaxed reach with a visibly bent elbow. Pell's shoulder is 8 cm lower, so the same panel is 0.21 m above her shoulder: she needs 0.52 m of arm, which is 90% of hers. The shorter character has to reach farther with a shorter arm. That inversion, not any single number, is the thing to carry into your own comparison.
Now run the transfer as rotations. Copying joint rotations onto Pell's shorter chain carries the palm 0.497 × (0.58 / 0.71) ≈ 0.41 m from her shoulder along the source's arm line. She ends up roughly 9 cm short along that line and about 14 cm from the contact point, low and short of the panel — with Mara's comfortable elbow bend intact, composing an arm around a contact it never makes.
Three routes come out of that, and each one leaves a different compromise:
Raw. Timing and phrasing are exactly the source's. The contact is wrong by a visible margin, and the arm's shape no longer means what it meant on Mara. Compromise: nothing is preserved except the motion's rhythm.
Corrected. Put a goal on the hand across the contact interval, and a goal on the planted foot across the same interval. The palm reaches the point. To do it, the elbow has to straighten from 70% toward 90% extension, and the shoulder has to lift or the torso has to lean more than the source did. Compromise: the contact is preserved and the performance is not — the hold now reads as strain.
Reanimated. Keep the contact point and the arrival frame, rebuild how the body arrives. Raise Pell's shoulder 8 cm and bring it 8 cm toward the panel — a rise onto the ball of the rear foot plus a lean in — and the required shoulder-to-contact distance drops to about 0.42 m, roughly 73% of her chain. The elbow keeps a bend close to Mara's, the strain moves into the legs and the stance, and the action reads as the action again. Compromise: it is a new piece of animation. The supplied motion becomes timing reference, nothing more.
Correct contact without freezing the rest of the body
A hand goal is a tool for a specific interval, not a property of the clip. Put it on the contact interval only. If the hand is driven to the panel for the entire animation, you have replaced the reach with a rail.
Then inspect what the rest of the body did to satisfy it. With a hand goal and no foot goal, the cheapest way for the solver to reach the point may be to move the pelvis — which is attached to the leg that is supposed to be planted. Watch the plant while the hand travels. A hand sitting exactly on the surface while the support foot has slid four backwards is not a preserved contact; it is two contacts, one fixed and one broken.
Check the contact in three parts, not one. Approach: does the hand arrive along a believable path, or does it cut a straight line to the goal? Hold: is the palm's angle against the surface right, or has the goal dragged a hand that is now flat, bent or rolled? Release: does the hand leave the way a hand leaves, or does the goal's end frame snap it away? Then check the shoulder, the pelvis and the support foot for compensating motion across all three, because a correction that solves the arm usually pays for it somewhere. Two contacts interact — hand and foot — and if both matter, both need goals over the same interval. If you start one interval before the other, the correction borrows from the plant to pay for the hand.
None of this makes the result natural. An IK goal proves that an end effector reached a point. It proves nothing about whether the reach looks like something a body would do.
Reanimate when proportion changes what the effort means
The criterion is not how much work correction took. It is whether the transferred action still says what it said. Mara flipping a switch with a bent elbow and Pell flipping the same switch with a locked-out arm are two different performances. One is routine; one is a stretch. If the beat in your pitch is "she does it without thinking," the corrected version contradicts the beat while technically satisfying the contact, and no amount of solver tuning will repair that.
So ask which read you need. If the character's effort is the point, the corrected version may be the better one — at which point make the strain deliberate, keyed and shaped, rather than left as the residue of a goal. If the contact is merely a piece of staging the beat sits on, and the effort change is incidental, correct it and move on.
Reanimation costs real time — new keys, re-derived timing against the same arrival frame, a different approach that has to work with the set as built. It also changes the leg contacts, which means the whole silhouette of the shot changes and you should say so rather than presenting it as a fix. What it buys is a character whose body is doing something it can actually do.
Judge the export, not the preview
Epic's retargeting overview notes that preview settings can differ from exported behaviour, root handling among them, which is enough to make the preview a working view rather than a receipt. Reopen the exported clip, repeat the same contact review with the same camera, and compare it against the raw version you saved. Check the frames around the arrival, the first and last frames, and whether the root motion you expected is present or baked away.
Record the engine build and the export settings alongside the clip. If you re-run the transfer on a different build later, you need to know which run produced what you showed, and a file with no build note cannot tell you.
Then decide with the exported action in front of you. If the raw transfer's only failure is contact geometry, and the corrected pose still reads as the performance you want, correct it and keep everything for comparison. If the target's proportions change what the effort means, animate the action and use the sample for its timing. Both are defensible. Neither is defensible from a screenshot of the arrival frame, or from a note in the file that the retarget worked — because the only thing worth proving, for a pitch test, is that the performance survived the transfer, and the only evidence of that is the exported clip meeting a contact point that never moved.
Frequently asked questions
What is actually being transferred when a supplied animation is retargeted?
The contact and the events around it: the fixed object, the plant foot at a point on the floor, the arrival frame, the hold, and the release, plus relationships such as weight, torso action, and gaze. The clip can play while the contact fails, which is why the arrival screenshot is not enough.
Why keep the uncorrected raw transfer?
It is the only honest baseline. It shows whether a failure is a constant offset—a correspondence problem—or a drift or miss that appears during the reach or at contact—a contact or performance problem. If you correct first and save nothing else, you cannot tell whether the corrected version is good or merely better than a version you can no longer see.
What should be checked before tuning anything?
Map whole joint chains, not bone-for-bone skeletons. Check where the arm and leg chains start and end, whether the clavicle or toe is included, the retarget root's orientation, the baseline pose on both characters, and whether unit scales differ. Epic's overview and walkthrough use vocabulary that may not match the build, so reconcile them against what is open.
How should the three routes be compared fairly?
Keep the contact point, arrival frame, review camera, and reviewed intervals constant. Raw changes nothing. Corrected may add goals or solvers and change joint rotations but must keep the same plant and foot contacts. Reanimated may rebuild the approach but must still put the palm on the same point at the same frame.
What should decide the final route, and where should you evaluate it?
If the raw transfer's only failure is contact geometry and the corrected pose still reads as the performance you want, correct it and keep everything for comparison. If the target's proportions change what the effort means, animate the action and use the sample for timing. Decide with the exported clip, not the preview, because preview settings can differ from exported behaviour, root handling among them. Record the engine build and export settings.