Add a physics engine (rapier3d) #110

New issue

Open

opened 2026-10-20 23:40:37 +00:00 by sermuns · 0 comments

sermuns commented

2026-10-20 23:40:37 +00:00

Owner

The campus has no physics. Everything that moves is moved by hand: the player
walks with a per-building 2-D wall-distance predicate
(collision::can_walk_to), the vehicles are a bespoke speed-and-heading model
(vehicle::step_vehicle), and the things on the ground — benches, bins,
bollards, seesaws — are drawn but never collided. Every roll, swing,
tumble, tip, fall over and bounce is a special case somebody has to
write by hand, and none of them compose.

Rapier is the obvious candidate: a pure-Rust physics engine with no renderer, a
first-class WASM story, and the joint, character-controller and ray-cast
vehicle APIs that the missing features all need. This issue is the spine; the
features it unlocks are the issues that link back to it.

What it would buy

A real vehicle: wheels, suspension, grip and rollover, instead of the
hand-rolled speed-and-heading model. #111
Objects that can be thrown, kicked and rolled: a football, a boule, a disc.
#112
Playground toys as mechanisms: a seesaw on a hinge, a swing on a pendulum.
#113
A body that can fall over instead of a body that slides along walls. #114
Doors that swing, and a building you can walk through properly. #115
Buoyancy: swimming, and things that float. #116

The costs, honestly

It contradicts a written decision. ARCHITECTURE.md says collision "is not
a swept capsule" and is a per-building 2-D wall-distance predicate, because
that keeps a walker on the right floor and never snaps a jump onto the slab
above. Rapier's character controller works the other way: one shape, cast
against the world. Getting the campus's storeys right means handing Rapier
the same floor-aware geometry the predicate has today — as one collider per
slab, or as collision groups — which is a bake change, not only a code
change.
Two vector vocabularies. The project has one: glam, with Vec2/Vec3
and the Enu/WorldPos newtypes. Rapier speaks nalgebra. The conversion
lives at the edge of the physics module, but it exists, and every crossing
pays for it.
The multiplayer split gets a third case. Players are client-authoritative
and vehicles are server-authoritative today. A physics world both must agree
on is something else: either the server owns every body and the 20 Hz pose
carries the state, or each client simulates its own and the two drift.
Determinism is opt-in. Rapier is deterministic on one machine by default.
Across machines it wants the enhanced-determinism feature, which cannot be
combined with simd8.
WASM gets no threads. Rapier builds for wasm32-unknown-unknown and its
own CI tests that target, but rayon falls back to one thread there, so the
web build pays single-threaded cost.

Where it would live

Rapier is pure Rust and has no rendering dependency, so it belongs in
liuverse-world beside collision, vehicle and player — the crate whose
whole job is "the simulation, without the renderer", and whose tests are pure
in-memory. It goes into [workspace.dependencies] like everything else.

The static world is the hard part. Rapier wants colliders, and the campus is
baked as triangle soups with footprints and rings. Two ways in:

Trimesh colliders from the baked geometry. Honest, and enormous: the whole
campus as one trimesh is a lot of triangles for a BVH to chew on, and it
collides the interiors with the exteriors.
Analytic colliders from the bake. One cuboid per floor slab, one per
interior wall ribbon, a height field for the terrain, a trimesh for the
roofs. The bake already knows how to produce all four. It is also what
makes the storeys work: a slab is a collider, so climbing a stair is a
slope and jumping under the next floor still cannot snap onto it.

The second is the one that keeps the current behaviour, so it is the one to try
first. It grows the bake a collision layer, which is a format change and
therefore a rebake.

The smaller step, if this is too much

parry3d is the same author's collision library without the dynamics solver.
It gives ray casts, shape casts and overlap queries — enough for picking and
for a proper camera — at a fraction of the weight and with no authority
questions. It is the right first move if the goal is picking and not
simulation.

Related

#111 the vehicles on real wheels
#112 the thrown and rolled bodies
#113 the playground mechanisms
#114 the ragdoll
#115 the doors
#116 the water
parry3d is the same author as rapier3d, one layer down.
ARCHITECTURE.md's collision and vehicle decisions are the ones that change.

The campus has no physics. Everything that moves is moved by hand: the player walks with a per-building 2-D wall-distance predicate (`collision::can_walk_to`), the vehicles are a bespoke speed-and-heading model (`vehicle::step_vehicle`), and the things on the ground — benches, bins, bollards, seesaws — are drawn but never collided. Every *roll*, *swing*, *tumble*, *tip*, *fall over* and *bounce* is a special case somebody has to write by hand, and none of them compose. Rapier is the obvious candidate: a pure-Rust physics engine with no renderer, a first-class WASM story, and the joint, character-controller and ray-cast vehicle APIs that the missing features all need. This issue is the spine; the features it unlocks are the issues that link back to it. **What it would buy** - A real vehicle: wheels, suspension, grip and rollover, instead of the hand-rolled speed-and-heading model. #111 - Objects that can be thrown, kicked and rolled: a football, a boule, a disc. #112 - Playground toys as mechanisms: a seesaw on a hinge, a swing on a pendulum. #113 - A body that can fall over instead of a body that slides along walls. #114 - Doors that swing, and a building you can walk through properly. #115 - Buoyancy: swimming, and things that float. #116 **The costs, honestly** - *It contradicts a written decision.* ARCHITECTURE.md says collision "is not a swept capsule" and is a per-building 2-D wall-distance predicate, because that keeps a walker on the right floor and never snaps a jump onto the slab above. Rapier's character controller works the other way: one shape, cast against the world. Getting the campus's storeys right means handing Rapier the same floor-aware geometry the predicate has today — as one collider per slab, or as collision groups — which is a bake change, not only a code change. - *Two vector vocabularies.* The project has one: `glam`, with `Vec2`/`Vec3` and the `Enu`/`WorldPos` newtypes. Rapier speaks `nalgebra`. The conversion lives at the edge of the physics module, but it exists, and every crossing pays for it. - *The multiplayer split gets a third case.* Players are client-authoritative and vehicles are server-authoritative today. A physics world both must agree on is something else: either the server owns every body and the 20 Hz pose carries the state, or each client simulates its own and the two drift. - *Determinism is opt-in.* Rapier is deterministic on one machine by default. Across machines it wants the `enhanced-determinism` feature, which cannot be combined with `simd8`. - *WASM gets no threads.* Rapier builds for `wasm32-unknown-unknown` and its own CI tests that target, but rayon falls back to one thread there, so the web build pays single-threaded cost. **Where it would live** Rapier is pure Rust and has no rendering dependency, so it belongs in `liuverse-world` beside `collision`, `vehicle` and `player` — the crate whose whole job is "the simulation, without the renderer", and whose tests are pure in-memory. It goes into `[workspace.dependencies]` like everything else. The static world is the hard part. Rapier wants colliders, and the campus is baked as triangle soups with footprints and rings. Two ways in: - *Trimesh colliders from the baked geometry.* Honest, and enormous: the whole campus as one trimesh is a lot of triangles for a BVH to chew on, and it collides the interiors with the exteriors. - *Analytic colliders from the bake.* One cuboid per floor slab, one per interior wall ribbon, a height field for the terrain, a trimesh for the roofs. The bake already knows how to produce all four. It is also what makes the storeys work: a slab is a collider, so climbing a stair is a slope and jumping under the next floor still cannot snap onto it. The second is the one that keeps the current behaviour, so it is the one to try first. It grows the bake a collision layer, which is a format change and therefore a rebake. **The smaller step, if this is too much** `parry3d` is the same author's collision library without the dynamics solver. It gives ray casts, shape casts and overlap queries — enough for picking and for a proper camera — at a fraction of the weight and with no authority questions. It is the right first move if the goal is *picking* and not *simulation*. **Related** - #111 the vehicles on real wheels - #112 the thrown and rolled bodies - #113 the playground mechanisms - #114 the ragdoll - #115 the doors - #116 the water - `parry3d` is the same author as `rapier3d`, one layer down. - ARCHITECTURE.md's collision and vehicle decisions are the ones that change.

sermuns added the

feature request

controversial

hard

labels

2026-10-20 23:40:37 +00:00

sermuns referenced this issue

2026-10-20 23:41:19 +00:00

Give the vehicles real wheels #111

sermuns referenced this issue

2026-10-20 23:41:20 +00:00

Throw and roll real objects #112