Implementation:Google deepmind Mujoco MJWarp Collision GJK

Knowledge Sources	Google_deepmind_Mujoco MJWarp Documentation
Domains	Physics_Simulation, GPU_Computing, NVIDIA_Warp
Last Updated	2026-02-15 04:00 GMT

Overview

MJWarp Collision GJK implements the Gilbert-Johnson-Keerthi distance algorithm, Expanding Polytope Algorithm, and Continuous Collision Detection for convex geometry on the GPU.

Description

This is the core convex collision detection module implementing GJK for distance computation between convex shapes, EPA for penetration depth calculation, and CCD for continuous collision detection. It defines Warp structs for GJKResult, Polytope, and SupportPoint, and provides support function evaluation for all MuJoCo geometry types. The module also implements the multicontact algorithm for generating multiple contact points from face-face and edge-face interactions using polygon clipping.

Usage

Used by the convex narrowphase module when geometry pairs require GJK-based distance computation or CCD. The gjk, ccd, multicontact, and support functions are the primary interfaces called from collision_convex.

Code Reference

Source Location

• Repository: Google_deepmind_Mujoco
• File: mjx/mujoco/mjx/third_party/mujoco_warp/_src/collision_gjk.py
• Lines: 1-2360

Key Functions

# Warp structs
@wp.struct
class GJKResult:
    dist: float; x1: wp.vec3; x2: wp.vec3; dim: int; ...

@wp.struct
class Polytope:
    status: int; vert: wp.array; face: wp.array; ...

@wp.struct
class SupportPoint:
    ...

# Core algorithms
def gjk(...)             # GJK distance algorithm
def ccd(...)             # Continuous collision detection
def multicontact(...)    # Multi-contact point generation via polygon clipping
def support(geom, geomtype, dir) -> SupportPoint  # Support function

# EPA (Expanding Polytope Algorithm)
def _epa(...)            # EPA penetration depth
def _epa_witness(...)    # EPA witness points

# Simplex sub-distance routines
def _S1D(s1, s2) -> wp.vec2
def _S2D(s1, s2, s3) -> wp.vec3
def _S3D(s1, s2, s3, s4) -> wp.vec4

Import

from mujoco.mjx.third_party.mujoco_warp._src.collision_gjk import gjk
from mujoco.mjx.third_party.mujoco_warp._src.collision_gjk import ccd
from mujoco.mjx.third_party.mujoco_warp._src.collision_gjk import multicontact
from mujoco.mjx.third_party.mujoco_warp._src.collision_gjk import support

I/O Contract

Inputs

Name	Type	Required	Description
geom1	Geom	Yes	First convex geometry with position, rotation, size, and mesh data
geom2	Geom	Yes	Second convex geometry
geomtype1	int	Yes	GeomType enum for first geometry
geomtype2	int	Yes	GeomType enum for second geometry

Outputs

Name	Type	Description
GJKResult	struct	Distance, witness points (x1, x2), simplex dimension and vertices
contacts	array	Contact positions, normals, and penetration depths from multicontact

Related Pages

• Environment:Google_deepmind_Mujoco_MJX_Warp_CUDA_Environment