Added Jolt physics example

[mrdoob]

Description

Just found out about Jolt which seems to be faster and more reliable than Rapier.

Made a JoltPhysics.js wrapper for it:

https://raw.githack.com/mrdoob/three.js/physics/examples/physics_jolt_instancing.html

Looks pretty good! Although the boxes seem to be overlapping 🤔

/fyi @jrouwe

[jrouwe]

To me it looks like there is a mismatch between the visual box size and the physics box size (although I don't spot the code bug), a simple shape like this should not penetrate.

[mrdoob]

Hmm... I'm going to have to investigate this further...

[DennisSmolek]

What kind of FPS were you getting?
I can't tell if it's just my computer but it was around 30fps.
I did a similar build for a react-three-jolt library and was staying around 60 until 800 or so but I was only using boxes and no shadows...

[Mugen87]

I've compared the demos physics_ammo_instancing, physics_rapier_instancing and the new physics_jolt_instancing on two devices. A mac mini with a M2 Pro runs all demos with 60 FPS. A Pixel 4a however does not:

• physics_ammo_instancing: 7 FPS
• physics_rapier_instancing: 30 FPS
• physics_jolt_instancing: 20 FPS

So it seems rapier is the fastest.

I did a similar build for a react-three-jolt library and was staying around 60 until 800 or so but I was only using boxes and no shadows...

We can only make a comparison if the example uses the exact same setup like the three examples mentioned above. Otherwise a performance comparison isn't valid. If you can share a live example with for testing, we might have a chance to find out why you see a performance difference between physics_jolt_instancing and a react-three-jolt demo.

[ycw]

boxes seem to be overlapping 🤔

box collision shape's convex radius (0.05) is too large for the visual cubes(side=0.075) 🤔

three.js/examples/jsm/physics/JoltPhysics.js

Line 21 in defb9fe

return new Jolt.BoxShape( new Jolt.Vec3( sx, sy, sz ), 0.05, null );

fyi, its less pronounced with smaller convex radius, e.g. 5% of min halfextent:

[Mugen87]

@ycw I've added your approach to the branch. It looks indeed much better!

[ycw]

nice

[mrdoob]

@jrouwe What do you think about this solution?

[jrouwe]

Given that the example uses boxes of 0.075, a convex radius of 0.05 is indeed too big (the debug version of Jolt will assert because it is bigger than half the size of the box). So this solution is better as it reduces the convex radius to 0.00375.

The default allowed penetration in Jolt is 0.02 (PhysicsSettings.mPenetrationSlop) which means the boxes can still sink into each other by this amount. Also, this will not help performance as the allowed penetration is bigger than the convex radius which causes a more expensive code path to trigger more often. You can update the physics settings through PhysicsSystem.SetPhysicsSettings.

[jrouwe]

@Mugen87

• physics_ammo_instancing: 7 FPS
• physics_rapier_instancing: 30 FPS
• physics_jolt_instancing: 20 FPS

So it seems rapier is the fastest.

There's at least one big difference between the Rapier and Jolt tests. Rapier is configured to use the delta time from the browser as timestep:

three.js/examples/jsm/physics/RapierPhysics.js

Lines 166 to 167 in 360c0e9

	world.timestep = clock.getDelta();
	world.step();

while Jolt is configured to try to not make steps bigger than 1/60th of a second:

three.js/examples/jsm/physics/JoltPhysics.js

Lines 213 to 222 in 360c0e9

	let deltaTime = clock.getDelta();
	// Don't go below 30 Hz to prevent spiral of death
	deltaTime = Math.min( deltaTime, 1.0 / 30.0 );
	// When running below 55 Hz, do 2 steps instead of 1
	const numSteps = deltaTime > 1.0 / 55.0 ? 2 : 1;
	// Step the physics world
	jolt.Step( deltaTime, numSteps );

So when the frame rate is below 55 fps (as is the case in your test), Jolt is configured to do 2 simulation steps while Rapier is configured to do 1 (if I understand the Rapier code correctly).

[mrdoob]

We should probably change it so both of them use the same approach.

@jrouwe What approach would you think would be better for someone starting? (better = causes less headaches)

[DennisSmolek]

For more basic/straightforward approach I'd just use the deltaTime. Drop the numSteps and just do one frame.

[jrouwe]

The danger of using deltaTime directly is that if the browser has a hiccup and deltaTime is suddenly a second, then the simulation will explode (this will happen on any physics engine that I know of). Removing numSteps is fine, but I would at least make sure that deltaTime is never bigger than 1/60th (although for simple scenes you may get away with 1/30th too).