Gummy Bears POC

Procedural gummy bears with real soft-body physics: one smooth-union SDF defines the bear once, and that definition produces the visible surface, the simulated particles, the limb clusters and the skin weights. Custom solver, no physics middleware.

Controls

Drag a bear to grab a paw and stretch it, release to let it snap back
Flick to throw a bear across the table
Space pours another bear, up to a bag of 16
R resets the bag
The panel sets flavour, wobble and candy memory, and reads out stretch, bounces and count

What You Are Looking At

95 particles and 5 clusters per bear, solved by cluster shape matching at 480 substeps per second
3 102 surface vertices skinned on the CPU each frame, faired with Taubin lambda/mu pairs in delta space
Translucent candy shading with Beer-Lambert absorption, plus coloured contact shadows that are depth-tested
Plasticity as a constitutive model: a stretched bear keeps part of the deformation and slowly recovers

WebGL 2 and a GPU help, but the simulation itself is plain TypeScript on the CPU: 1.06 ms per frame for one bear, 4.60 ms for a bag of four. The demo is served from this site, so it works offline once the page has loaded. If the frame stays blank,open it in its own tab.

How This Build Got Here

The demo is a build artefact of the argon proof of concept, rebuilt with a site-relative base and copied into site/public/projects/gummy-bear/app/ by./scripts/automation/sync-gummy-bear.sh. The full engineering write-up, including the headless vision-critique loop that tuned it, is onHow the argon POC Was Built. Source: private repository on the owner’s Forgejo (not published).