Water-tight rigid–soft contact

An SDF-based contact path so a deforming body cannot tunnel through a rigid shape between its vertices, with full integration into the VBD solver. Newton · SolverVBD · NVIDIA L40 · 2026-06.

3 passes

particle + soft-edge + soft-face contact generation

FACE + EDGE

contacts with no vertex required at the contact

flag off ⇒ identical

byte-for-byte legacy behaviour when disabled

two-way

soft side and rigid-body reaction in VBD

Results — gripping a soft mesh by a vertex-free feature

A parallel-jaw gripper (cube pads) closes on a soft mesh positioned so that no mesh vertex lies between the jaws — only the interior of a face or an edge does — then lifts and holds. Left = water-tight OFF reproduces main: the legacy per-particle path only tests vertices, finds none in the jaws, so the mesh slips through and drops. Right = water-tight ON: the mesh is caught by its face / edge and held in the air.

Soft triangle — gripped by its FACE

The triangle's three vertices are outside the jaws; only the face crosses the pinch. OFF: it falls. ON: a water-tight face contact holds it.

Soft 1×1 grid — gripped by its interior DIAGONAL EDGE

The grid's four corners are outside the jaws; the diagonal edge crosses the pinch. OFF: it falls. ON: a water-tight edge contact holds it (the sheet drapes over the grip).

What changed

The feature is purely additive and gated by enable_water_tight_rigid_soft_contact; with the flag off the contact set and solver results are byte-for-byte identical to before.

1 · Contact generation (SDF back-end)

Three additive passes per collision: the legacy particle pass, a soft-edge pass (1-D golden-section search along each unique edge), and a soft-face pass (Frank–Wolfe over the triangle's barycentric simplex), each minimising the rigid shape's signed-distance field — following Macklin et al. 2020.
Analytic primitives use their closed-form SDF; meshes use a volume (texture) SDF, provisioned at finalize() for participating meshes that lack one.
Records pack into a standalone soft counter laid out as [particle, edge, face]. The internal soft_contact_particle is renamed soft_contact_primitive (now a particle or soft-triangle id, by range), with new soft_contact_kind / soft_contact_barycentric fields.

2 · VBD integration

Soft side. The particle-contact kernel gains a second section over the edge/face range. The contact point is the barycentric combination x = Σ bᵢ·vᵢ of the soft triangle's corners, so the penalty force and Hessian distribute as bᵢ·F and bᵢ²·H to each vertex (an edge leaves one weight at zero).
Rigid side (two-way). A per-slot kernel applies the equal-and-opposite reaction at the rigid contact point to dynamic bodies, so the gripper feels the grasp. Both sides reuse the existing penalty law and penetration > 0 gate.

3 · Tests & examples

Detection tests (SDF optimisers vs. a brute-force grid, edge/face passes, end-to-end no false-positive/negative, flag-off backward-compat) and VBD tests (per-vertex barycentric force/Hessian distribution, the two grips above as held-in-the-air regressions). Green on CPU and CUDA.
Two runnable examples: vbd_gripper_soft_triangle and vbd_gripper_soft_grid (the videos above; toggle enable_water_tight to reproduce the OFF / main case).

Gotcha worth noting: an analytic box face gives an ambiguous SDF normal when a soft triangle straddles it (the closest feature flips between the bottom and a side face), so the two-way grip tests use convex shapes (cube / sphere) where the normal is unique and stable.

SDF memory footprint

How much memory each shape's SDF actually uses, measured through the real finalize() path at the default settings (64³ max resolution, UINT16-quantised narrow band, 8³ subgrids). Analytic primitives store nothing; only meshes allocate a volume SDF.

Shape	SDF representation	Stored memory
Box, Sphere, Capsule, Cylinder, Cone, Ellipsoid, Plane	closed-form analytic (evaluated, not stored)	0
Mesh — bunny (12,200 tris)	texture volume SDF	≈ 0.48 MB
Mesh — bear (3,968 tris)	texture volume SDF	≈ 0.30 MB

All analytic shapes report shape_sdf_index = -1 under water-tight — nothing is allocated; the contact path evaluates their closed-form SDF.

Mesh breakdown — the narrow-band subgrid dominates

Component	bunny	bear
coarse texture (far field, float32)	2.5 KB (9×9×8)	1.3 KB (9×6×6)
subgrid texture (narrow band, uint16)	488 KB (63³)	307 KB (54³)
indirection slots (uint32)	1.8 KB	0.8 KB
total	493 KB	310 KB

Sparse narrow-band storage: cost scales with surface extent, not triangle count — the 4k-tri bear is smaller than the 12k-tri bunny.
Allocated once per unique source mesh (deduplicated) and shared across instances — N bunnies still cost ~0.48 MB.
Knobs: UINT16 → float32 roughly doubles the subgrid; max_resolution scales the narrow band by ~the square of the ratio.

SDF accuracy

Texture-SDF value vs. the exact mesh signed distance (winding-number query) at 300k points sampled within ±6 voxels of the surface (the contact-relevant band). The error is resolution-limited, so it is reported both absolutely and as a fraction of one voxel (≈ object extent / 64); the meshes are at their native USD scale.

Mesh	voxel	MAE	RMSE	p99	max
bunny	26.7 mm	1.02 mm (3.8% vox)	1.86 mm (6.9%)	7.8 mm (29%)	22 mm (83%)
bear	152.7 mm	7.30 mm (4.8% vox)	14.6 mm (9.6%)	51 mm (33%)	206 mm (135%)

Mean error is ~4–5% of a voxel (≈ 0.04–0.06% of the bounding-box diagonal) — both meshes land in the same relative band, confirming the accuracy is set by resolution, not the model.
Error is dominated by the finite 64³ resolution (trilinear interpolation), not the UINT16 quantisation (whose step over the narrow band is sub-micron at this scale).
Worst case is ≈ one voxel, at thin / high-curvature features a trilinear grid can't resolve below voxel size.
Accuracy trades against memory: doubling max_resolution roughly halves the error and ~8×'s the subgrid.