[GodotPhysics] Raycasts don't reliably collide with HeightMapShape3D

[celeste-sinead]

Godot version

v4.0.beta.custom_build.01ae26d31

System information

Pop!_OS 22.04, 5.19.0-76051900-generic, Intel(R) UHD Graphics (CML GT2)

Issue description

When casting rays from the camera in the direction of the mouse position, I found that in some camera positions and orientations, the result of the result of PhysicsDirectSpaceState3D.intersect_ray is empty where it ought to have found a collision.

When this occurs, there is generally a contiguous region of the heightmap through which rays cast from a given camera position won't produce a result. Different camera transforms produce different non-colliding regions (and some camera transforms result in no such regions). When I originally encountered this issue with non-flat heightmaps, I found non-colliding regions could occur anywhere in the heightmap, not necessarily touching edges, in non-rectangular shapes, and often multiple such regions would exist.

In the simplified reproduction project I've only observed rectangular non-colliding regions, usually at the +X edge but sometimes at +Z.

Steps to reproduce

I haven't isolated exact conditions for when this occurs. I've observed it to be sensitive to the following:

• heightmap size: I've only been able to reproduce this with heightmaps of 20x20 and larger
• raycast length: I've only been able to reproduce this with rays >20 length
• camera transform (ray origin / direction): changes which regions of the heightmap don't collide, and in some cases prevent the issue. Directions closer to -Y seem much less likely to reproduce the issue.
• heightmap shape: the flat heightmap in the repro project only seems to produce rectangular non-colliding regions at the edge of the heightmap, but for non-flat (specifically, simplex noise) maps I've seen multiple not-necessarily rectangular regions not necessarily touching edges.

The reproduction project sets a camera transform where the issue is apparent (and will cast rays to mouse position and position a sphere at the intersect location when detected). It also includes a script that casts an example ray that ought to collide but doesn't, and also draws a line corresponding to that ray to demonstrate visually that it ought to have collided.

Minimal reproduction project

heightmap_raycast_repro.zip

[Calinou]

Can you reproduce this on 3.5.1 (with both Bullet and GodotPhysics)?

[celeste-sinead]

This reproduces on 3.5.1 with GodotPhysics but not with Bullet.

And also on 4.0 beta 4 (e675154), for what it's worth.

[ciqven]

I can confirm this is still occurring on Godot 4.0 beta 16, with a flat (map_data nothing but zeroes) HeightMapShape3D of size 64x64.

What works

• Using a BoxShape3D works.
• Using create_trimesh_collision() on the MeshInstance3D representing the terrain works.
• Enabling Debug -> Visible Collision Shapes shows a correct looking wireframe for HeightMapShape3D.

What doesn't

• When using a HeightMapShape3D, manual raycasting with PhysicsDirectSpaceState3D.intersect_ray() returns an empty Dictionary for some regions of the shape.
• When using a HeightMapShape3D, CollisionObject3D.input_event() does not receive any InputEvents for the same regions as previous point.

[dkaste]

I can also reproduce this (v4.1.dev.custom_build [5860b02])

In my case, I am intersecting rays projected from the mouse cursor with the heightmap shape. Only very specific rays (~1px mouse position difference) return an empty dictionary.

I have a hunch that this may happen at the borders between the collision shape triangles as seen with Visible Collision Shapes enabled.