Dennis Gustafsson

@voxagonlabs • 84,244 subscribers

Game designer and programmer. Founder of @tuxedolabs. Teardown @teardowngame, Smash Hit, PinOut, Does not Commute, Sprinkle, Granny Smith, Meqon and Dresscode.

Live Cam

Dennis Gustafsson

Videos

Watch Anya Live

Some physically driven animation for your Wednesday afternoon. After months of trying to get RL to look nice, I've switched over to good old procedural animation.

The spaghetti stress test. Each batch is about 800 bodies, connected with shape-preserving ball joints.

Experimenting with an in-game build mode. A lot of people have been requesting that. I can see why :)

Joints in the multi-threaded solver. While simpler to implement in theory, multiple violating constraints are much more sensitive to ordering, making them challenging for graph coloring.

It's a strange job, but look at that stability!

The new voxel format is finally starting to come together. It uses a brickmap that enables large, sparse objects, unique properties for each voxel (no palette) and really fast scene updates.

Time to put those deformable voxels to work. Here is a very quick softbody test. There are some artefacts and missing motion vectors but there's potential!

Colorful deformable voxel sausages with self-collision. If you have no idea what this is useful for, that makes two of us.

Some work-in-progress third person camera action and procedural driving animations.

I finally got graph coloring working properly in the solver with good performance. Now large piles can be solved on multiple threads in deterministic order. It really helps with stability. Big thanks to Erin Catto pointing me in the right direction!

I've been trying to implement a parallel rigid body solver. It's a tricky problem in physics to solve a huge pile of objects on multiple threads. I've tried graph coloring and other means of splitting up the problem, but so far a locking primitive per body is the most efficient..

Experimenting with a new way of doing stochastic transparency. It still uses a checkerboard pattern, but allows for both variable transparency, tinted and frosted glass by recombining transparent and opaque pixels in a post-pass. Looks pretty promising so far!

I've done some graphics research lately and been tinkering with a new voxel renderer. It's pretty different from the one in Teardown, focusing heavily on GI. There are no explicit light sources in this scene (except for the sun), just emissive voxels.

It's getting better, but progress is pretty slow. The reward here is for staying upright while matching a rest post. This one has about 65k parameters trained with PPO for half an hour.

Baby steps towards physically based animation with reinforcement learning...

Reinforcement learning experiments continues. Still a bit janky, but it is clearly learning something. This is about an hour into training.