Issues

• Sphere and capsule contacts use [1, 0, 0] (or float32 noise) as the normal when the closest points coincide; MuJoCo C uses the cross product of the geom z-axes

  google-deepmind/mujoco_warp#1961 · Filter by author dundysm opened on Nov 13, 2026

• Capsule-on-plane contacts use a fixed tangent when the capsule is steeper than 60°, unlike C MuJoCo

  google-deepmind/mujoco_warp#1949 · Filter by author YannBerthelot opened on Nov 12, 2026

• Complete contact and constraint ordering behind DeterminismType.ALL

  google-deepmind/mujoco_warp#1925 · Filter by author huster-wgm opened on Nov 8, 2026
  1

• Joint limit sensors read tendon limit rows with the same id

  google-deepmind/mujoco_warp#1910 · Filter by author Nas01010101 opened on Nov 6, 2026

• Joint and tendon limits create one row when both sides are within the margin

  google-deepmind/mujoco_warp#1909 · Filter by author Nas01010101 opened on Nov 6, 2026

• Performance: set_const INERT0 solves M one right-hand side at a time and scales poorly at large nworld

  google-deepmind/mujoco_warp#1895 · Filter by author NeoZng opened on Nov 5, 2026

• njmax=0: Euler and implicit integrators advance with qacc = 0, so the model never moves

  google-deepmind/mujoco_warp#1887 · Filter by author fl0fischer opened on Nov 4, 2026
  3

• A contact from a single clipped multicontact point gets the GJK/EPA distance instead of its clipped distance (MuJoCo C uses the clipped one)

  google-deepmind/mujoco_warp#1886 · Filter by author lulu-loopp opened on Nov 3, 2026
  3

• Box-box contact normals differ from MuJoCo C by up to the tilt: box-box goes to GJK/EPA multicontact unless NATIVECCD is disabled, MuJoCo C always uses mjc_BoxBox

  google-deepmind/mujoco_warp#1885 · Filter by author lulu-loopp opened on Nov 3, 2026

• Box-box contacts flipped for a thin box resting on a slab (normal reversed, depth = summed thickness)

  google-deepmind/mujoco_warp#1883 · Filter by author Codnay opened on Nov 3, 2026

• GJK returns a positive distance for a 1-3 mm box-box penetration, so EPA is skipped and no contact is written

  google-deepmind/mujoco_warp#1851 · Filter by author mhmalmir opened on Oct 27, 2026
  1

• Direct MyoSuite CPU/CUDA muscle rollouts drift at the wrist with contacts disabled

  google-deepmind/mujoco_warp#1835 · Filter by author M-Colley opened on Oct 22, 2026
  1

• Proposal: 2D flex collisions with static heightfields

  google-deepmind/mujoco_warp#1831 · Filter by author cola1917 opened on Oct 22, 2026

• Constant-height heightfield box contacts differ from equivalent plane manifold

  google-deepmind/mujoco_warp#1826 · Filter by author huster-wgm opened on Oct 22, 2026
  1

• Heightfield-box witness correction produces downward penetration at step edges

  google-deepmind/mujoco_warp#1824 · Filter by author huster-wgm opened on Oct 22, 2026
  2

• opt.disableactuator (actuatorgroupdisable) is silently ignored

  google-deepmind/mujoco_warp#1806 · Filter by author dundysm opened on Oct 20, 2026

• accelerometer, framelinacc and frameangacc report -gravity on static and mocap bodies

  google-deepmind/mujoco_warp#1804 · Filter by author dundysm opened on Oct 20, 2026
  1

• Feature request: per-world fixed tendon coefficients (batched wrap_prm)

  google-deepmind/mujoco_warp#1799 · Filter by author balmung08 opened on Oct 18, 2026
  

• Exactly aligned convex-mesh stacks never settle: CCD drops a penetrating face-face pair that MuJoCo C detects

  google-deepmind/mujoco_warp#1798 · Filter by author almightychang opened on Oct 18, 2026
  1

• set_const_fixed: every world writes the same geom_aabb row when geom_size is batched but geom_aabb isn't

  google-deepmind/mujoco_warp#1795 · Filter by author Codnay opened on Oct 17, 2026
  1

• CUDA EPA near-degenerate face produces a contact position 410 m away in cylinder–box collision

  google-deepmind/mujoco_warp#1782 · Filter by author coverMoon opened on Oct 13, 2026
  2

• Collision sensors use contact-buffer slots, and an overflow drops real contacts

  google-deepmind/mujoco_warp#1759 · Filter by author ayonga opened on Oct 7, 2026

• Distance sensor between two capsules reports the cutoff unless they penetrate

  google-deepmind/mujoco_warp#1758 · Filter by author ayonga opened on Oct 7, 2026

• camlight produces degenerate target camera and light frames

  google-deepmind/mujoco_warp#1743 · Filter by author kavyabhand opened on Oct 4, 2026
  1

• GJK: in float32 the barycentric coordinates of a small or thin simplex do not sum to 1 (capsule depths on the sphere/capsule path off by up to 12.8 mm, penetrating contacts missed)

  google-deepmind/mujoco_warp#1740 · Filter by author M-Colley opened on Oct 3, 2026