Single precision: EPA face normals tilt on exactly aligned face pairs since mjMINDIST2/3 were lowered to 1e-17

Abierto
#3,914 1 comentario 0 reacciones 0 asignados Ver en GitHub

Los mantenedores suelen responder en 6 días

@almightychang ya está trabajando en esto.

Desde el 17/11/2026.

  • #3915 de @almightychang — abierto

Evaluación

Dificultad
4/5
Tiempo estimado
3-5 días
Aptitud para principiantes
35/100
Tipo de issue
Error
Claridad
Bien especificado
Estado de actividad
Estancado
Stack tecnológico
cpp
Área
backend

Línea de trabajo

Start with the single precision distance tolerances and EPA face normal calculations in engine_collision_gjk.c, then compare behavior with the reproduction model and program in the issue. The reported candidate fix should preserve shallow penetration handling while restoring aligned contact normals and depths; verify single precision behavior and unchanged double precision results. Pull request #3915 is already open against this issue.

Escrito por el modelo de indexación a partir del texto del issue.

Descripción

Intro

I contribute to MuJoCo Warp, which runs MuJoCo's GJK/EPA in float32 with the constants of the mjUSESINGLE build, so I compare against MuJoCo C built with mjUSESINGLE.

My setup

MuJoCo C at main (aa9a0eaef), built with CMake and -DmjUSESINGLE in CMAKE_C_FLAGS and CMAKE_CXX_FLAGS. Linux x86_64, GCC. Double precision builds of the same commit for comparison.

What's happening? What did you expect?

Since 3bbf510d4 ("Improve NativeCCD numerical precision"), a stack of thin box-shaped mesh plates resting exactly aligned on a mesh slab no longer settles in the single precision build. The four corner contacts of a face pair, which should have the same depth and the normal (0, 0, 1), get normals tilted by 3e-4 to 5e-4 rad and depths that differ by up to about 100 µm (for example -172, -209, -286, -249 µm on the top pair at step 999). The plates rock, and CG needs 10 to 20 times more iterations.

Mean CG iterations over the second half of 2000 steps, and max z drift of any plate over that half:

plates (8 mm thick) double, main single, 3bbf510d4^ single, main
3 1.3, 0.000 mm 1.3, 0.000 mm 16.8, 0.018 mm
4 1.8, 0.001 mm 1.8, 0.001 mm 26.4, 0.074 mm
5 2.0, 0.001 mm 2.0, 0.001 mm 19.9, 0.156 mm

Within 3bbf510d4, the change comes from the single precision mjMINDIST2 and mjMINDIST3, lowered from 1e-10f to 1e-17f: setting them back to 1e-10f restores equal depths, the (0, 0, 1) normal and 1 to 2 iterations. Reverting them is not a fix, though, because the lower values are what made shallow penetration work in single precision: with 1e-10f, mj_geomDistance returns 0 for aligned boxes or mesh boxes overlapping by 2 to 12 µm.

I have a candidate fix in engine_collision_gjk.c that keeps both and leaves double precision results unchanged, and can send it as a PR.

Steps for reproduction
  1. Build MuJoCo with -DmjUSESINGLE.
  2. Compile the program below against it (with -DmjUSESINGLE) and run it on the model.
  3. Compare with a double precision build, or with single precision at 3bbf510d4^.
Minimal model for reproduction
<mujoco>
  <option timestep="0.002" solver="CG" cone="elliptic" impratio="10" integrator="implicitfast"
          iterations="100" ls_iterations="15" tolerance="1e-6"/>
  <asset>
    <mesh name="plate" vertex="-0.105 -0.148 -0.004  -0.105 -0.148 0.004  -0.105 0.148 -0.004  -0.105 0.148 0.004
                               0.105 -0.148 -0.004  0.105 -0.148 0.004  0.105 0.148 -0.004  0.105 0.148 0.004"/>
    <mesh name="slab" vertex="-0.3 -0.3 -0.02  -0.3 -0.3 0.02  -0.3 0.3 -0.02  -0.3 0.3 0.02
                              0.3 -0.3 -0.02  0.3 -0.3 0.02  0.3 0.3 -0.02  0.3 0.3 0.02"/>
  </asset>
  <worldbody>
    <body pos="0 0 0.02"><geom type="mesh" mesh="slab"/></body>
    <body pos="0 0 0.044"><freejoint/><geom type="mesh" mesh="plate" mass="0.05"/></body>
    <body pos="0 0 0.052"><freejoint/><geom type="mesh" mesh="plate" mass="0.05"/></body>
    <body pos="0 0 0.060"><freejoint/><geom type="mesh" mesh="plate" mass="0.05"/></body>
  </worldbody>
</mujoco>
Code required for reproduction
#include <math.h>
#include <stdio.h>
#include <mujoco/mujoco.h>

int main(int argc, char** argv) {
  char err[1000];
  mjModel* m = mj_loadXML(argv[1], NULL, err, sizeof(err));
  mjData* d = mj_makeData(m);
  double iters = 0, drift = 0, z0[3];
  for (int step = 0; step < 2000; step++) {
    mj_step(m, d);
    if (step == 999) {
      for (int c = 0; c < d->ncon; c++) {
        printf("geoms %d-%d dist %+.1f um normal (%.4f %.4f %.4f)\n", d->contact[c].geom[0], d->contact[c].geom[1],
               1e6 * d->contact[c].dist, d->contact[c].frame[0], d->contact[c].frame[1], d->contact[c].frame[2]);
      }
    }
    if (step >= 1000) {
      iters += d->solver_niter[0];
      for (int p = 0; p < 3; p++) {
        double z = d->qpos[7 * p + 2];
        if (step == 1000) z0[p] = z;
        drift = fmax(drift, fabs(z - z0[p]));
      }
    }
  }
  printf("mean CG iterations %.1f, max z drift %.3f mm\n", iters / 1000, 1e3 * drift);
  return 0;
}
Confirmations
  • I searched the latest documentation thoroughly before posting.
  • I searched previous Issues and Discussions, I am certain this has not been raised before.
Lenguaje dominante
C++
Estrellas
18.3k
Forks
2.2k
Merge medio
9 d 11 h
PR fusionados (30 d)
5

Preparar el entorno

Primeros pasos

  1. Lee el issue completo y luego la guía de contribución del proyecto.
  2. Comenta en el issue que vas a ocuparte — evita que dos personas hagan lo mismo.
  3. Haz un fork del repositorio y trabaja en una rama.
  4. Abre un pull request que haga referencia al número del issue.

Más de google-deepmind/mujoco

Todos los issues de google-deepmind/mujoco

Issues similares

Más issues de C++

Recibe los nuevos issues en tu correo

Un resumen breve de issues de GitHub para principiantes.