A worker walking home from the gold mine used to look exactly like a worker walking anywhere else. Now it carries a sack of gold on its back, a bundle of logs if it has been chopping, or a pack of quarried stone, and the load is gone the moment it is dropped off.
The request came straight after the gold mine started lighting up while workers are inside. That fixed the moment a worker disappears. Adam wanted the rest of the trip to read too: a visual cue that a worker is carrying something, and which resource it is. Ten ground workers across the factions meant one load had to sit convincingly on a broad dwarf, a small goblin, a hunched ratman, a skink with a tail and a slim elf.
One load, drawn twice
The game has a 2D renderer and a 3D one, and both read the same rule for when a load shows: the worker is carrying something, it is gold, lumber or stone, and it is not hidden inside a mine. Oil is left out on purpose. Only tankers carry it, and a ship has no back.
In 2D the load is a small sprite over the upper body, pushed away from the way the worker is facing. Facing the camera or sideways, it draws behind the body, so you see the top of the sack over the shoulders. Facing away, it draws over the body and you see all of it.
The 3D loads are models from Meshy, generated from those same reference images. The stone pack went through four versions of its reference image before one was approved, and that version has no shoulder straps. Straps would have to wrap ten different body shapes, and a strapless pack can simply sit flush against whatever back it is on.
A raw Meshy export of a prop is around 10,000 triangles with a 2048 pixel texture, for an object drawn a few dozen pixels tall. A small tool built on glTF-Transform and meshoptimizer simplifies each one to about 1,500 triangles, shrinks the texture to 512 pixels, and re-anchors the model so its base sits on the ground plane and the face that rests against a back sits at zero depth. The rest of the placement code assumes exactly that frame.
Why the load is not parented to a bone
The obvious way to attach a backpack is to make it a child of a spine bone. That breaks here, because a unit in the 3D renderer is not one mesh. It is a ladder of detail levels, the subject of the level of detail post, and only one is drawn at a time. A prop parented to one level's bone disappears the moment the ladder swaps to another.
So the load is a child of the unit's root, which already carries its position, fog visibility and everything else about the unit, and every frame it copies its placement from whichever detail level is active. On a skinned level that is the chest bone. On the coarsest static levels, which have no skeleton, it falls back to a fixed point at about half the model's height, behind the body.
The bone called Spine02 is the lower back
The first version anchored every load on the bone named Spine02. It is the obvious choice by name, and it was wrong on every worker. On a Meshy humanoid rig the chain runs from the hips to Spine02, then Spine01, then Spine, and the shoulders and neck hang off that last one. Spine02 is the small of the back.
Lined up in the lab, it was plain. Loads sat low, sank into the chunkier torsos, and stayed upright while the body under them bent, because they took the model's facing and nothing else. A load strapped to a back leans and twists with the chest.
The fix finds the chest by structure instead of by name: it is the bone the neck hangs off, which holds on any humanoid rig whatever its spine happens to be called. Bone axes are arbitrary between rigs, so the prop's orientation is not taken from the bone either. It is built from anatomy, in the bind pose, and stored as an offset from the chest bone's rotation. After that, every frame is one multiplication, and the load leans, twists and bobs exactly as the chest does.
import * as THREE from 'three';
// A torso frame from anatomy: up is lower spine to neck, right is shoulder to
// shoulder, and "out" (away from the back) is their cross product.
// Pass bind-pose positions, so every instance of a model gets the same offset.
export function torsoOffset(chestBindQuat, lowerSpine, neck, leftShoulder, rightShoulder) {
const up = new THREE.Vector3().subVectors(neck, lowerSpine).normalize();
const right = new THREE.Vector3().subVectors(rightShoulder, leftShoulder);
const out = new THREE.Vector3().crossVectors(right, up).normalize();
right.crossVectors(up, out).normalize();
const frame = new THREE.Quaternion().setFromRotationMatrix(new THREE.Matrix4().makeBasis(right, up, out));
return chestBindQuat.clone().invert().multiply(frame);
}
// Each frame: the chest's live rotation times the offset is the torso's frame.
export function torsoFrame(chestBone, offset, target = new THREE.Quaternion()) {
return chestBone.getWorldQuaternion(target).multiply(offset);
}Two details are easy to miss. Find the chest as the parent of the neck bone, not by guessing a spine name. And check which way "out" points against something that knows where the face is: Meshy rigs carry a bone just ahead of the head, and a mirrored rig with left and right swapped would otherwise put the load on the worker's chest. Deciding that from a live pose instead is fragile. One worker's idle animation bends almost double, and it had its load put inside its chest.
Measuring ten backs instead of guessing ten offsets
With the anchor right, the remaining question is how big each load should be and how far out from the chest it should sit. A dwarf's back is a barrel well behind its chest bone. A skink is thin but hunched. A ratman has a rounded hump. Six numbers per worker per load (size, out from the back, up, sideways, lean and spin) across ten workers and three loads is 180 numbers, and tuning them by eye in the carry lab was the first plan.
The lab now fits them instead. It takes each worker's real skinned mesh in the poses a load is mostly seen in, standing idle and at four points of the walk cycle, and measures the back in the same torso frame the load is placed in. Size comes from shoulder breadth, height from the space between the hips and the shoulders, and lean from the slope of the back across the load's height, so a hunched back gets a load that lies along it.
What counts as back took some deciding. Arms and legs are excluded, or a swinging arm in the walk cycle would shove every load a hand's width off the body. So is everything below the hips: Meshy rigs have no tail bones, so a ratman's or skink's tail is skinned to the hips and sticks straight out behind, and measured as body it pushed every load about a third of a body height off the back.
Loads that hovered, and loads that sank
The first auto-fit still looked wrong on the contact sheets, a grid of every worker wearing every load, from behind and in profile, idle and walking. Almost every load hovered slightly off its back. The fit had pushed each load out until it cleared the single furthest point of the back, so a flat pack touched a rounded back at one spot and stood proud everywhere else.
The second version matches the load's actual surface against the back's, patch by patch, and pushes each one in until the closest quarter of the contact sits just inside the body, by 1.2% of the worker's height. A cap stops any part going deeper than 6%, because without it a round sack disappeared into the slimmer backs. One more exception came from the elves: their long hair is skinned to the head bone, the fit ignored it, and loads sank into the hair. Anything skinned to the head now counts as back when it hangs below the neck.
All 30 worker and load fits were re-run and checked on the sheets from behind and in profile, standing and walking, before they shipped.
Small things that make a load belong to the unit
A load casts a shadow when the player has unit shadows on, since it is part of the unit. The three prop models load with the rest of a match's models, so the first worker to leave a mine never pops a load in a frame late. The gold sack also borrows the gold mine's glow: the same colour mask that lights the veins in the rock picks out the coins spilling from the sack, tuned so the burlap stays dull.
Fitting your own props to many characters
If one prop has to sit on many different bodies, do not tune an offset per character by eye. Measure instead. Put the prop in a frame built from the skeleton's anatomy, sample the skinned mesh in the two or three poses the prop is actually seen in, leave out limbs and anything below the hips, and fit the prop's surface against the body's rather than against a single point. Then photograph every combination from behind and in profile on one sheet, because a bad fit is invisible on the one character you happen to be looking at.
Every fit is stored as a fraction of the worker's height, so rescaling a worker keeps its load in place. A new worker added to the roster needs one run of the auto-fit and one look at its row on the sheet.
