MuJoCo Warp (MJWarp)#

MuJoCo Warp (MJWarp) is an implementation of MuJoCo written in Warp and optimized for NVIDIA hardware and parallel simulation. MJWarp lives in the google-deepmind/mujoco_warp GitHub repository.

MJWarp is developed and maintained as a joint effort by NVIDIA and Google DeepMind.

Tutorial notebook#

The MJWarp basics are covered in a tutorial [notebook] [open in colab].

When To Use MJWarp?#

High throughput#

The MuJoCo ecosystem offers multiple options for batched simulation.

mujoco.rollout: Python API for multi-threaded calls to mj_step on CPU. High throughput can be achieved with hardware that has fast cores and large thread counts, but overall performance of applications requiring frequent host<>device transfers (e.g., reinforcement learning with simulation on CPU and learning on GPU) may be bottlenecked by transfer overhead.
mjx.step: jax.vmap and jax.pmap enable multi-threaded and multi-device simulation with JAX on CPUs, GPUs, or TPUs.
mujoco_warp.step: Python API for multi-threaded and multi-device simulation with CUDA via Warp on NVIDIA GPUs. Improved scaling for contact-rich scenes compared to the MJX JAX implementation.

Low latency#

MJWarp is optimized for throughput: the total number of simulation steps per unit time whereas MuJoCo is optimized for latency: time for one simulation step. It is expected that a simulation step with MJWarp will be less performant than a step with MuJoCo for the same simulation.

As a result, MJWarp is well suited for applications where large numbers of samples are required, like reinforcement learning, while MuJoCo is likely more useful for real-time applications like online control (e.g., model predictive control) or interactive graphical interfaces (e.g., simulation-based teleoperation).

Complex scenes#

MJWarp scales better than MJX for scenes with many geoms or degrees of freedom, but not as well as MuJoCo. There may be significant performance degradation in MJWarp for scenes beyond 60 DoFs. Supporting these larger scenes is a high priority and progress is tracked in GitHub issues for: sparse Jacobians #88, block Cholesky factorization and solve #320, constraint islands #886, and sleeping islands #887.

Differentiability#

The dynamics API in MJX is automatically differentiable via JAX. We are considering whether to support this in MJWarp via Warp - if this feature is important to you, please chime in on this issue here.

Installation#

From PyPI:

pip install mujoco-warp

From source:

git clone https://github.com/google-deepmind/mujoco_warp.git
cd mujoco_warp
uv sync --all-extras

To make sure everything is working:

uv run pytest -n 8

Basic Usage#

Once installed, the package can be imported via import mujoco_warp as mjw. Structs, functions, and enums are available directly from the top-level mjw module.

Structs#

Before running MJWarp functions on an NVIDIA GPU, structs must be copied onto the device via mjw.put_model and mjw.make_data or mjw.put_data functions. Placing an mjModel on device yields an mjw.Model. Placing an mjData on device yields an mjw.Data.

mjm = mujoco.MjModel.from_xml_string("...")
mjd = mujoco.MjData(mjm)
m = mjw.put_model(mjm)
d = mjw.put_data(mjm, mjd)

These MJWarp variants mirror their MuJoCo counterparts but have a few key differences:

mjw.Model and mjw.Data contain Warp arrays that are copied onto device.
Some fields are missing from mjw.Model and mjw.Data for features that are unsupported.

Batch sizes#

MJWarp is optimized for parallel simulation. A batch of simulations can be specified with three parameters:

nworld: Number of worlds to simulate.
nconmax: Expected number of contacts per world. The maximum number of contacts for all worlds is nconmax * nworld.
naconmax: Alternative to nconmax, maximum number of contacts over all worlds. If nconmax and naconmax are both set then nconmax is ignored.
njmax: Maximum number of constraints per world.

Semantic difference for nconmax and njmax.

It is possible for the number of contacts per world to exceed nconmax if the total number of contacts for all worlds does not exceed nworld x nconmax. However, the number of constraints per world is strictly limited by njmax.

XML parsing

Values for nconmax and njmax are not parsed from size/nconmax and size/njmax (these parameters are deprecated). Values for these parameters must be provided to mjw.make_data or mjw.put_data.

Functions#

MuJoCo functions are exposed as MJWarp functions of the same name, but following PEP 8-compliant names. Most of the main simulation and some of the sub-components for forward simulation are available from the top-level mjw module.

Minimal example#

# Throw a ball at 100 different velocities.

import mujoco
import mujoco_warp as mjw
import warp as wp

_MJCF=r"""
<mujoco>
  <worldbody>
    <body>
      <freejoint/>
      <geom size=".15" mass="1" type="sphere"/>
    </body>
  </worldbody>
</mujoco>
"""

mjm = mujoco.MjModel.from_xml_string(_MJCF)
m = mjw.put_model(mjm)
d = mjw.make_data(mjm, nworld=100)

# initialize velocities
wp.copy(d.qvel, wp.array([[float(i) / 100, 0, 0, 0, 0, 0] for i in range(100)], dtype=float))

# simulate physics
mjw.step(m, d)

print(f'qpos:\n{d.qpos.numpy()}')

Command line scripts#

Benchmark an environment with testspeed

mjwarp-testspeed benchmark/humanoid/humanoid.xml

Interactive environment simulation with MJWarp

mjwarp-viewer benchmark/humanoid/humanoid.xml

Feature Parity#

MJWarp supports most of the main simulation features of MuJoCo, with a few exceptions. MJWarp will raise an exception if asked to copy to device an mjModel with field values referencing unsupported features.

The following features are not supported in MJWarp:

Category	Feature
Integrator	`IMPLICIT`, `IMPLICITFAST` not supported with fluid drag
Solver	`PGS`, `noslip`, islands
Fluid Model	Ellipsoid model
Sensors	`GEOMDIST`, `GEOMNORMAL`, `GEOMFROMTO`
Flex	`VERTCOLLIDE=false`, `INTERNAL=true`
Jacobian format	`SPARSE`
Option	contact override
Plugins	`All` except `SDF`
User parameters	`All`

Performance Tuning#

The following are considerations for optimizing the performance of MJWarp.

Graph capture#

MJWarp functions, for example mjw.step, often comprise a collection of kernel launches. Warp will launch these kernels individually if the function is called directly. To improve performance, especially if the function will be called multiple times, it is recommended to capture the operations that comprise the function as a CUDA graph

with wp.ScopedCapture() as capture:
  mjw.step(m, d)

The graph can then be launched or re-launched

wp.capture_launch(capture.graph)

and will typically be significantly faster compared to calling the function directly. Please see the Warp Graph API reference for details.

Batch sizes#

The maximum numbers of contacts and constraints, nconmax / naconmax and njmax respectively, are specified when creating mjw.Data with mjw.make_data or mjw.put_data. Memory and computation scales with the values of these parameters. For best performance, the values of these parameters should be set as small as possible while ensuring the simulation does not exceed these limits.

It is expected that good values for these limits will be environment specific. In practice, selecting good values typically involves trial-and-error. mjwarp-testspeed with the flag --measure_alloc for printing the number of contacts and constraints at each simulation step and interacting with the simulation via mjwarp-viewer and checking for overflow errors can both be useful techniques for iteratively testing values for these parameters.

Solver iterations#

MuJoCo’s default solver settings for the maximum numbers of solver iterations and linesearch iterations are expected to provide reasonable performance. Reducing MJWarp’s settings Option.iterations and/or Option.ls_iterations limits may improve performance and should be secondary considerations after tuning nconmax / naconmax and njmax.

Reducing these limits too much may prevent the constraint solver from converging and can lead to inaccurate or unstable simulation.

Impact on Performance: MJX (JAX) and MJWarp

In MJX these solver parameters are key for controlling simulation performance. With MJWarp, in contrast, once all worlds have converged the solver can early exit and avoid unnecessary computation. As a result, the values of these settings have comparatively less impact on performance.

Contact sensor matching#

Scenes that include contact sensors have a parameter that specifies the maximum number of matched contacts per sensor Option.contact_sensor_max_match. For best performance, the value of this parameter should be as small as possible while ensuring the simulation does not exceed the limit. Matched contacts that exceed this limit will be ignored.

The value of this parameter can be set directly, for example model.opt.contact_sensor_maxmatch = 16, or via an XML custom numeric field

<custom>
  <numeric name="contact_sensor_maxmatch" data="16"/>
</custom>

Similar to the maximum numbers of contacts and constraints, a good value for this setting is expected to be environment specific. mjwarp-testspeed and mjwarp-viewer may be useful for tuning the value of this parameter.

Parallel linesearch#

In addition to the constraint solver’s iterative linesearch, MJWarp provides a parallel linesearch routine that evaluates a set of step sizes in parallel and selects the best one. The step sizes are spaced logarithmically from Model.opt.ls_parallel_min_step to 1 and the number of step sizes to evaluate is set via Model.opt.ls_iterations.

In some cases the parallel routine may provide improved performance compared to the constraint solver’s default iterative linesearch.

To enable this routine set Model.opt.ls_parallel=True or add a custom numeric field to the XML

<custom>
  <numeric name="ls_parallel" data="1"/>
</custom>

Experimental feature

The parallel linesearch is currently an experimental feature.

Memory#

Simulation throughput is often limited by memory requirements for large numbers of worlds. Considerations for optimizing memory utilization include:

CCD colliders require more memory than primitive colliders, see MuJoCo’s pair-wise colliders table for information about colliders.
multiccd requires more memory than CCD.
CCD memory requirements scale linearly with Option.ccd_iterations.
A scene with at least one mesh geom and using multiccd will have memory requirements that scale linearly with the maximum number of vertices per face and with the maximum number of edges per vertex, computed over all meshes.

testspeed provides the flag --memory for reporting a simulation’s total memory utilization and information about mjw.Model and mjw.Data fields that require significant memory. Memory allocated inline, including for CCD and the constraint solver, can also be significant and is reported as Other memory.

Maximum number of contacts per collider

Some MJWarp colliders have a different maximum number of contacts compared to MuJoCo:

PLANE<>MESH: 4 versus 3
HFieldCCD: 4 versus mjMAXCONPAIR

Sparsity

Sparse Jacobians can enable significant memory savings. Updates for this feature are tracked in GitHub issue #88.

The mjw.make_data or mjw.put_data argument nccdmax / naccdmax can be set to a value less than nconmax / naconmax in order to reduce the memory requirements for CCD. The value for this parameter should be the maximum number of contacts generated by a CCD collider, per world or for all worlds, respectively. For example, a batched simulation with 10 worlds that generates 80 total contacts with per-collider contacts: mesh-mesh: 30 (CCD), ellipsoid-ellipsoid: 10 (CCD), and sphere-sphere: 40 (primitive) should set nconmax / naconmax to at least 8 / 80 (may require more for broadphase) and nccdmax / naccdmax to 3 / 30.

Batched `Model` Fields#

To enable batched simulation with different model parameter values, many mjw.Model fields have a leading batch dimension. By default, the leading dimension is 1 (i.e., field.shape[0] == 1) and the same value(s) will be applied to all worlds. It is possible to override one of these fields with a wp.array that has a leading dimension greater than one. This field will be indexed with a modulo operation of the world id and batch dimension: field[worldid % field.shape[0]].

Graph capture

The field array should be overridden prior to graph capture (i.e., wp.ScopedCapture) since the update will not be applied to an existing graph.

# override shape and values
m.dof_damping = wp.array([[0.1], [0.2]], dtype=float)

with wp.ScopedCapture() as capture:
  mjw.step(m, d)

It is possible to override the field shape and set the field values after graph capture

# override shape
m.dof_damping = wp.empty((2, 1), dtype=float)

with wp.ScopedCapture() as capture:
  mjw.step(m, d)

# set batched values
dof_damping = wp.array([[0.1], [0.2]], dtype=float)
wp.copy(m.dof_damping, dof_damping)  # m.dof = dof_damping will not update the captured graph

Modifying fields#

The recommended workflow for modifying an mjModel field is to first modify the corresponding mjSpec and then compile to create a new mjModel with the updated field. However, compilation currently requires a host call: 1 call per new field instance, i.e., nworld host calls for nworld instances.

Certain fields are safe to modify directly without compilation, enabling on-device updates. Please see mjModel changes for details about specific fields. Additionally, GitHub issue 893 tracks adding on-device updates for a subset of fields.

Per-world meshes#

Per-world meshes enable heterogeneous worlds where different worlds simulate different meshes. The workflow is:

Create an mjSpec with all mesh assets and the maximum number of geom slots needed across variants.
Compile each variant by mutating the spec and calling spec.compile().
Compile a base model and create mjw.Model from it.
Override the relevant mjw.Model fields with per-world arrays built from the compiled variants.

Example 1 — Geom-level randomization (1 body, 1 geom, 2 mesh assets):

The base scene includes all mesh assets. The geom references one mesh (mesh_a); a second mesh (mesh_b) is available for per-world substitution.

<mujoco>
  <asset>
    <mesh name="mesh_a" vertex="0 0 0 1 0 0 0 1 0 0 0 1"/>
    <mesh name="mesh_b" vertex="0 0 0 2 0 0 0 2 0 0 0 2"/>
  </asset>
  <worldbody>
    <body pos="0 0 1">
      <freejoint/>
      <geom name="obj" type="mesh" mesh="mesh_a"/>
    </body>
  </worldbody>
</mujoco>

nworld = 4

# base spec: 1 body with 1 mesh geom, all mesh assets
spec = mujoco.MjSpec()
mesh_a = spec.add_mesh()
mesh_a.name = "mesh_a"
mesh_a.uservert = [0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0, 1]

mesh_b = spec.add_mesh()
mesh_b.name = "mesh_b"
mesh_b.uservert = [0, 0, 0, 2, 0, 0, 0, 2, 0, 0, 0, 2]

body = spec.worldbody.add_body()
body.pos = [0, 0, 1]
body.add_freejoint()
geom = body.add_geom()
geom.name = "obj"
geom.type = mujoco.mjtGeom.mjGEOM_MESH
geom.meshname = "mesh_a"

# compile each variant
geom.meshname = "mesh_a"
mjm_a = spec.compile()
geom.meshname = "mesh_b"
mjm_b = spec.compile()

# restore and compile base
geom.meshname = "mesh_a"
mjm = spec.compile()

m = mjw.put_model(mjm)
d = mjw.make_data(mjm, nworld=nworld)

# build per-world arrays: worlds 0-1 use mesh_a, worlds 2-3 use mesh_b
geom_id = mujoco.mj_name2id(mjm, mujoco.mjtObj.mjOBJ_GEOM, "obj")
variants = [mjm_a, mjm_b]
assignment = [0, 0, 1, 1]  # variant index per world

# build per-world arrays
dataid = np.tile(mjm.geom_dataid, (nworld, 1))
geom_size = np.zeros((nworld, mjm.ngeom, 3))
geom_aabb = np.zeros((nworld, mjm.ngeom, 2, 3))
geom_rbound = np.zeros((nworld, mjm.ngeom))
geom_pos = np.zeros((nworld, mjm.ngeom, 3))
body_mass = np.zeros((nworld, mjm.nbody))
body_subtreemass = np.zeros((nworld, mjm.nbody))
body_inertia = np.zeros((nworld, mjm.nbody, 3))
body_invweight0 = np.zeros((nworld, mjm.nbody, 2))
body_ipos = np.zeros((nworld, mjm.nbody, 3))
body_iquat = np.zeros((nworld, mjm.nbody, 4))

for w in range(nworld):
  ref = variants[assignment[w]]
  dataid[w, geom_id] = ref.geom_dataid[geom_id]
  geom_size[w] = ref.geom_size
  geom_aabb[w] = ref.geom_aabb.reshape(mjm.ngeom, 2, 3)
  geom_rbound[w] = ref.geom_rbound
  geom_pos[w] = ref.geom_pos
  body_mass[w] = ref.body_mass
  body_subtreemass[w] = ref.body_subtreemass
  body_inertia[w] = ref.body_inertia
  body_invweight0[w] = ref.body_invweight0
  body_ipos[w] = ref.body_ipos
  body_iquat[w] = ref.body_iquat

m.geom_dataid = wp.array(dataid, dtype=int)
m.geom_size = wp.array(geom_size, dtype=wp.vec3)
m.geom_aabb = wp.array(geom_aabb, dtype=wp.vec3)
m.geom_rbound = wp.array(geom_rbound, dtype=float)
m.geom_pos = wp.array(geom_pos, dtype=wp.vec3)
m.body_mass = wp.array(body_mass, dtype=float)
m.body_subtreemass = wp.array(body_subtreemass, dtype=float)
m.body_inertia = wp.array(body_inertia, dtype=wp.vec3)
m.body_invweight0 = wp.array(body_invweight0, dtype=wp.vec2)
m.body_ipos = wp.array(body_ipos, dtype=wp.vec3)
m.body_iquat = wp.array(body_iquat, dtype=wp.quat)

Example 2 — Body-level randomization (1 body, 1 or 2 geoms, 3 mesh assets):

Maximum geom count

For body-level randomization, the base mjModel provided to mjw.put_model should specify the maximum number of geoms required across all variants. Geom slots that are unused in a particular variant can be disabled (e.g., contype=0, conaffinity=0, dataid=-1), but they should still be present as part of the body in the base model.

<mujoco>
  <asset>
    <mesh name="mA" vertex="0 0 0 1 0 0 0 1 0 0 0 1"/>
    <mesh name="mB" vertex="0 0 0 2 0 0 0 2 0 0 0 2"/>
    <mesh name="mC" vertex="0 0 0 3 0 0 0 3 0 0 0 3"/>
  </asset>
  <worldbody>
    <body name="obj" pos="0 0 1">
      <freejoint/>
      <geom name="obj_0" type="mesh" mesh="mA"/>
      <geom name="obj_1" size=".001" contype="0" conaffinity="0" mass="0"/>
    </body>
  </worldbody>
</mujoco>

nworld = 6

# base spec: body with 2 geom slots (max across variants), all mesh assets
spec = mujoco.MjSpec()
for name, scale in [("mA", 1), ("mB", 2), ("mC", 3)]:
  mesh = spec.add_mesh()
  mesh.name = name
  mesh.uservert = [0, 0, 0, scale, 0, 0, 0, scale, 0, 0, 0, scale]

body = spec.worldbody.add_body()
body.name = "obj"
body.pos = [0, 0, 1]
body.add_freejoint()

g0 = body.add_geom()
g0.name = "obj_0"
g0.type = mujoco.mjtGeom.mjGEOM_MESH
g0.meshname = "mA"

# null geom slot: disabled collision, no mesh
g1 = body.add_geom()
g1.name = "obj_1"
g1.size = [0.001, 0, 0]
g1.contype = 0
g1.conaffinity = 0
g1.mass = 0

# variant A: 1 geom (mesh mA), g1 stays null
mjm_a = spec.compile()

# variant B: 2 geoms (mesh mB + mC)
g0.meshname = "mB"
g1.type = mujoco.mjtGeom.mjGEOM_MESH
g1.meshname = "mC"
g1.contype = 1
g1.conaffinity = 1
mjm_b = spec.compile()

# restore base and compile
g0.meshname = "mA"
g1.type = mujoco.mjtGeom.mjGEOM_SPHERE
g1.contype = 0
g1.conaffinity = 0
mjm = spec.compile()

m = mjw.put_model(mjm)
d = mjw.make_data(mjm, nworld=nworld)

# worlds 0-2: variant A (1 active geom), worlds 3-5: variant B (2 active geoms)
variants = [mjm_a, mjm_b]
assignment = [0, 0, 0, 1, 1, 1]

geom0_id = mujoco.mj_name2id(mjm, mujoco.mjtObj.mjOBJ_GEOM, "obj_0")
geom1_id = mujoco.mj_name2id(mjm, mujoco.mjtObj.mjOBJ_GEOM, "obj_1")
body_id = mjm.geom_bodyid[geom0_id]

# build per-world arrays
dataid = np.tile(mjm.geom_dataid, (nworld, 1))
geom_size = np.zeros((nworld, mjm.ngeom, 3))
geom_rbound = np.zeros((nworld, mjm.ngeom))
geom_aabb = np.zeros((nworld, mjm.ngeom, 2, 3))
geom_pos = np.zeros((nworld, mjm.ngeom, 3))
body_mass = np.zeros((nworld, mjm.nbody))
body_subtreemass = np.zeros((nworld, mjm.nbody))
body_inertia = np.zeros((nworld, mjm.nbody, 3))
body_invweight0 = np.zeros((nworld, mjm.nbody, 2))
body_ipos = np.zeros((nworld, mjm.nbody, 3))
body_iquat = np.zeros((nworld, mjm.nbody, 4))

for w in range(nworld):
  ref = variants[assignment[w]]
  dataid[w] = ref.geom_dataid
  # disable unused geom slot for variant A
  if assignment[w] == 0:
    dataid[w, geom1_id] = -1
  geom_size[w] = ref.geom_size
  geom_rbound[w] = ref.geom_rbound
  geom_aabb[w] = ref.geom_aabb.reshape(mjm.ngeom, 2, 3)
  geom_pos[w] = ref.geom_pos
  body_mass[w] = ref.body_mass
  body_subtreemass[w] = ref.body_subtreemass
  body_inertia[w] = ref.body_inertia
  body_invweight0[w] = ref.body_invweight0
  body_ipos[w] = ref.body_ipos
  body_iquat[w] = ref.body_iquat

m.geom_dataid = wp.array(dataid, dtype=int)
m.geom_size = wp.array(geom_size, dtype=wp.vec3)
m.geom_rbound = wp.array(geom_rbound, dtype=float)
m.geom_aabb = wp.array(geom_aabb, dtype=wp.vec3)
m.geom_pos = wp.array(geom_pos, dtype=wp.vec3)
m.body_mass = wp.array(body_mass, dtype=float)
m.body_subtreemass = wp.array(body_subtreemass, dtype=float)
m.body_inertia = wp.array(body_inertia, dtype=wp.vec3)
m.body_invweight0 = wp.array(body_invweight0, dtype=wp.vec2)
m.body_ipos = wp.array(body_ipos, dtype=wp.vec3)
m.body_iquat = wp.array(body_iquat, dtype=wp.quat)

Batched fields — fields that must be overridden for per-world meshes:

Field	dtype	Shape
`geom_dataid`	`int`	`(nworld, ngeom)`
`geom_size`	`wp.vec3`	`(nworld, ngeom)`
`geom_aabb`	`wp.vec3`	`(nworld, ngeom, 2)`
`geom_rbound`	`float`	`(nworld, ngeom)`
`geom_pos`	`wp.vec3`	`(nworld, ngeom)`
`body_mass`	`float`	`(nworld, nbody)`
`body_subtreemass`	`float`	`(nworld, nbody)`
`body_inertia`	`wp.vec3`	`(nworld, nbody)`
`body_invweight0`	`wp.vec2`	`(nworld, nbody)`
`body_ipos`	`wp.vec3`	`(nworld, nbody)`
`body_iquat`	`wp.quat`	`(nworld, nbody)`

Batch Rendering#

MJWarp provides a batch renderer for high-throughput ray tracing built on Warp’s accelerated BVHs for rendering worlds with multiple cameras in parallel.

Key features:

Mesh rendering with textures: BVH-accelerated mesh rendering with full texture support.
Heightfield rendering: Optimized rendering for heightfields.
Flex rendering: Render 2D and 3D flex objects.
Lighting and shadows: Dynamic lighting with configurable shadows; domain randomizable: light_active, light_type, light_castshadow, light_xpos, light_xdir.
Heterogeneous multi-camera: Multiple cameras per world and each camera can have a different resolution (cam_resolution), field of view (cam_fovy, cam_sensorsize, cam_intrinsic), and output mode (cam_output).
Domain Randomization: Per-world mjw.Model fields (see Batched Model Fields above): geom_matid, geom_size, geom_rgba, mat_texid, mat_texrepeat, mat_rgba.
BVH-accelerated ray/rays API: Ray casting: Accelerated mjw.ray, mjw.rays, and rangefinder sensors via Warp’s BVHs.

Basic Usage#

Rendering or raycasting requires a mjw.RenderContext which contains BVH structures, rendering specific fields, and output buffers.

rc = mjw.create_render_context(
    mjm,
    nworld=1,
    cam_res=(256, 256),           # Override camera resolution (or per-camera list)
    render_rgb=True,              # Enable RGB output (or per-camera list)
    render_depth=True,            # Enable depth output (or per-camera list)
    use_textures=True,            # Apply material textures
    use_shadows=False,            # Enable shadow casting (slower)
    enabled_geom_groups=[0, 1],   # Only render geoms in groups 0 and 1
    cam_active=[True, False],     # Selectively enable/disable cameras
    flex_render_smooth=True,      # Smooth shading for soft bodies
)

Each mjw.RenderContext parameter can be applied globally or per camera. Additionally, values for mjw.RenderContext parameters can be parsed from XML:

<camera name="front_camera" pos="3 0 2" xyaxes="0 1 0 -0.6 0 0.8" resolution="64 64" output="rgb depth"/>

or set via mjSpec for camera customization.

To render, first call mjw.refit_bvh to update the BVH trees, followed by mjw.render to write to output buffers.

mjw.refit_bvh(m, d, rc)
mjw.render(m, d, rc)

The output buffers contain stacked pixels for all cameras with shape (nworld, npixel) and RGB data is packed into one uint32 variable. RenderContext.rgb_adr and RenderContext.depth_adr provide per-camera indexing. For convenience, mjw.get_rgb and mjw.get_depth return processed and reshaped RGB and depth data for a given camera batched for all worlds.

nworld = 1
cam_index = 0
resolution = rc.cam_res.numpy()[cam_index]
rgb_data = wp.zeros((nworld, resolution[1], resolution[0]), dtype=wp.vec3)
mjw.get_rgb(rc, rgb_data=rgb_data, cam_id=cam_index)

A complete example can be found in the MJWarp tutorial [notebook] [open in colab].

Benchmarks#

Rendering can be benchmarked using testspeed:

mjwarp-testspeed benchmarks/primitives.xml --function=render

For benchmark results across a variety of scenes, see the released benchmarks.

Notes#

Meshes: Rendering computation scales with mesh complexity, specifically the number of vertices and faces. A primitive is expected to have better performance (i.e., higher throughput) compared to a similar-sized mesh or heightfield.
Flex: Currently limited to 2D and 3D flex objects. Performance is expected to improve as this feature is further developed.
Scaling: Rendering scales linearly with resolution (total pixel count) and camera count.

Frequently Asked Questions#

Learning frameworks#

Does MJWarp work with JAX?

Yes. MJWarp is interoperable with JAX. Please see the Warp Interoperability documentation for details.

Additionally, MJX provides a JAX API for a subset of MJWarp’s API. The implementation is specified with impl='warp'.

Does MJWarp work with PyTorch?

Yes. MJWarp is interoperable with PyTorch. Please see the Warp Interoperability documentation for details.

How to train policies with MJWarp physics?

For examples that train policies with MJWarp physics, please see:

Isaac Lab: Train via Newton API.
mjlab: Train directly with MJWarp using PyTorch.
MuJoCo Playground: Train via MJX API.

Features#

Is MJWarp differentiable?

No. MJWarp is not currently differentiable via Warp’s automatic differentiation functionality. Updates from the team related to enabling automatic differentiation for MJWarp are tracked in this GitHub issue.

Does MJWarp work with multiple GPUs?

Yes. Warp’s wp.ScopedDevice enables multi-GPU computation

# create a graph for each device
graph = {}
for device in wp.get_cuda_devices():
  with wp.ScopedDevice(device):
    m = mjw.put_model(mjm)
    d = mjw.make_data(mjm)
    with wp.ScopedCapture(device) as capture:
      mjw.step(m, d)
    graph[device] = capture.graph

# launch a graph on each device
for device in wp.get_cuda_devices():
  wp.capture_launch(graph[device])

Please see the Warp documentation for details and mjlab distributed training for a reinforcement learning example.

Is MJWarp on GPU deterministic?

No. There may be ordering or small numerical differences between results computed by different executions of the same code. This is characteristic of non-deterministic atomic operations on GPU. Set device to CPU with wp.set_device("cpu") for deterministic results.

Developments for deterministic results on GPU are tracked in this GitHub issue.

How are orientations represented?

Orientations are represented as unit quaternions and follow MuJoCo’s conventions: w, x, y, z or scalar, vector.

wp.quaternion

MJWarp utilizes Warp’s built-in type wp.quaternion. Importantly however, MJWarp does not utilize Warp’s x, y, z, w quaternion convention or operations and instead implements quaternion routines that follow MuJoCo’s conventions. Please see math.py for the implementations.

Does MJWarp have a named access API / bind?

No. Updates for this feature are tracked in this GitHub issue.

Why are contacts reported when there are no collisions?

1 contact will be reported for each unique geom pair that contributes to any collision sensor, even if this geom pair is not in collision. Unlike MuJoCo or MJX where collision sensors make separate calls to collision routines while computing sensor data, MJWarp computes and stores the data for these sensors in contacts while running its main collision pipeline.

Contact sensors will report the correct information for contacts affecting the physics.

Why are Jacobians always dense?

Sparse Jacobians are not currently implemented and Data fields: ten_J, actuator_moment, flexedge_J, and efc.J are always represented as dense matrices. Support for sparse Jacobians is tracked in GitHub issue #88.

Why do some arrays have different shapes compared to mjModel or mjData?

By default for batched simulation, many mjw.Data fields having a leading batch dimension of size Data.nworld. Some mjw.Model fields having a leading batch dimension with size 1, indicating that this field can be overridden with an array of batched parameters for domain randomization.

Additionally, certain fields including Model.qM, Data.efc.J, and Data.efc.D are padded to enable fast loading on GPU.

Why are numerical results from MJWarp and MuJoCo different?

MJWarp utilizes float <https://nvidia.github.io/warp/modules/functions.html#warp.float32>`__s in contrast to MuJoCo's default double representation for :ref:`mjtNum. Solver settings, including iterations, collision detection, and small friction values may be sensitive to differences in floating point representation.

If you encounter unexpected results, including NaNs, please open a GitHub issue.

Why is inertia matrix qM sparsity not consistent with MuJoCo / MJX?

mjtJacobian semantics

MuJoCo’s inertia matrix is always sparse and mjtJacobian affects constraint Jacobians and related quantities
MJWarp’s (and MJX’s) constraint Jacobian is always dense and mjtJacobian is repurposed to affect the inertia matrix that can be represented as dense or sparse

The automatic sparsity threshold utilized by MJWarp for AUTO is optimized for GPU and set to nv > 32, unlike MuJoCo and MJX which use nv >= 60. Dense DENSE and sparse SPARSE settings are consistent with MuJoCo and MJX.

This feature is likely to change in the future.

How to fix simulation runtime warnings?

Warnings are provided when memory requirements exceed existing allocations during simulation:

nconmax / njmax: The maximum number of contacts / constraints has been exceeded. Increase the value of the setting by updating the relevant argument to mjw.make_data or mjw.put_data.
mjw.Option.ccd_iterations: The convex collision detection algorithm has exceeded the maximum number of iterations. Increase the value of this setting in the XML / mjSpec / mjModel. Importantly, this change must be made to the mjModel instance that is provided to mjw.put_model and mjw.make_data / mjw.put_data.
mjw.Option.contact_sensor_maxmatch: The maximum number of contact matches for a contact sensor’s matching criteria has been exceeded. Increase the value of this MJWarp-only setting m.opt.contact_sensor_maxmatch. Alternatively, refactor the contact sensor matching criteria, for example if the 2 geoms of interest are known, specify geom1 and geom2.
height field collision overflow: The number of potential contacts generated by a height field exceeds mjMAXCONPAIR and some contacts are ignored. To resolve this warning, reduce the height field resolution or reduce the size of the geom interacting with the height field.

Compilation#

How can compilation time be improved?

Limit the number of unique colliders that require the general convex collision pipeline. These colliders are listed as _CONVEX_COLLISION_PAIRS in collision_convex.py. Improvements to the compilation time for the pipeline are tracked in this GitHub issue.

Why are the physics not working as expected after upgrading MJWarp?

The Warp cache may be incompatible with the current code and should be cleared as part of the debugging process. This can be accomplished by deleting the directory ~/.cache/warp or via Python

import warp as wp
wp.clear_kernel_cache()

Is it possible to compile MJWarp ahead of time instead of at runtime?

Yes. Please see Warp’s Ahead-of-Time Compilation Workflows documentation for details.

Differences from MuJoCo#

This section notes differences between MJWarp and MuJoCo.

Warmstart#

If warmstarts are not disabled, the MJWarp solver warmstart always initializes the acceleration with qacc_warmstart. In contrast, MuJoCo performs a comparison between qacc_smooth and qacc_warmstart to determine which one is utilized for the initialization.

Inertia matrix factorization#

When using dense computation, MJWarp’s factorization of the inertia matrix qLD is computed with Warp’s L'L Cholesky factorization wp.tile_cholesky and the result is not expected to match MuJoCo’s corresponding field because a different reverse-mode L'DL routine mj_factorM is utilized.

Options#

mjw.Option fields correspond to their mjOption counterparts with the following exceptions:

impratio is stored as its inverse square root impratio_invsqrt.
The constraint solver setting tolerance is clamped to a minimum value of 1e-6.
Contact override parameters o_margin, o_solref, o_solimp, and o_friction are not available.

disableflags has the following differences:

mjDSBL_MIDPHASE is not available.
mjDSBL_AUTORESET is not available.
mjDSBL_NATIVECCD changes the default box-box collider from CCD to a primitive collider.
mjDSBL_ISLAND is not currently available. Constraint island discovery is tracked in GitHub issue #886.

enableflags has the following differences:

mjENBL_OVERRIDE is not available.
mjENBL_FWDINV is not available.
Constraint island sleeping enabled via mjENBL_ISLAND is not currently available. This feature is tracked in GitHub issues #886 and #887.

Additional MJWarp-only options are available:

ls_parallel: use parallel linesearch with the constraint solver
ls_parallel_min_step: minimum step size for the parallel linesearch
broadphase: type of broadphase algorithm (mjw.BroadphaseType)
broadphase_filter: type of filtering utilized by broadphase (mjw.BroadphaseFilter)
graph_conditional: use CUDA graph conditional
run_collision_detection: use collision detection routine
contact_sensor_maxmatch: maximum number of contacts for contact sensor matching criteria

Fluid model

Modifying fluid model parameters: density, viscosity, or wind may require updating Model.has_fluid.

Graph capture

A new graph capture may be necessary after modifying an mjw.Option field in order for the updated setting to take effect.

SDF plugins#

SDF collisions support plugins. The following example for plugin/sdf/bowl.xml illustrates how to implement the SDF plugin implementation in bowl.cc:

import mujoco_warp as mjw
import warp as wp

# distance function
@wp.func
def bowl(p: wp.vec3, attr: wp.vec3) -> float:
  """Signed distance function for a bowl shape.

  attr[0] = height
  attr[1] = radius
  attr[2] = thickness
  """
  height = attr[0]
  radius = attr[1]
  thick = attr[2]
  width = wp.sqrt(radius * radius - height * height)

  # q = (norm_xy(p), p.z)
  q0 = wp.sqrt(p[0] * p[0] + p[1] * p[1])
  q1 = p[2]

  # qdiff = q - (width, height)
  qdiff0 = q0 - width
  qdiff1 = q1 - height

  if height * q0 < width * q1:
    dist = wp.sqrt(qdiff0 * qdiff0 + qdiff1 * qdiff1)
  else:
    q_norm = wp.sqrt(q0 * q0 + q1 * q1)
    dist = wp.abs(q_norm - radius)

  return dist - thick


# gradient of distance function
@wp.func
def bowl_sdf_grad(p: wp.vec3, attr: wp.vec3) -> wp.vec3:
  """Gradient of bowl SDF via finite differences."""
  eps = float(1e-6)
  f0 = bowl(p, attr)

  px = wp.vec3(p[0] + eps, p[1], p[2])
  py = wp.vec3(p[0], p[1] + eps, p[2])
  pz = wp.vec3(p[0], p[1], p[2] + eps)

  grad = wp.vec3(
    (bowl(px, attr) - f0) / eps,
    (bowl(py, attr) - f0) / eps,
    (bowl(pz, attr) - f0) / eps,
  )
  return grad


# register the bowl SDF plugin
@wp.func
def user_sdf(p: wp.vec3, attr: wp.vec3, sdf_type: int) -> float:
  return bowl(p, attr)


@wp.func
def user_sdf_grad(p: wp.vec3, attr: wp.vec3, sdf_type: int) -> wp.vec3:
  return bowl_sdf_grad(p, attr)


# override the module-level hooks
mjw._src.collision_sdf.user_sdf = user_sdf
mjw._src.collision_sdf.user_sdf_grad = user_sdf_grad

Physics callbacks#

MuJoCo provides global physics callbacks that allow users to inject custom logic into the simulation pipeline. MJWarp supports a similar mechanism, but callbacks are Python functions set per-model on the mjw.Model instance via Model.callback rather than as global function pointers.

The following callbacks are available:

Callback	Description
`control`	Custom control laws, writes to `Data.ctrl`
`passive`	Custom passive forces, writes to `Data.qfrc_passive`
`act_dyn`	Custom actuator dynamics, writes to `Data.act_dot`
`act_gain`	Custom actuator gains, writes to `Data.actuator_force`
`act_bias`	Custom actuator biases, writes to `Data.actuator_force`
`sensor`	Custom sensors, writes to `Data.sensordata`; receives an additional `stage` argument
`contactfilter`	Custom contact filtering, writes to `Data.contact`

import mujoco
import mujoco_warp as mjw
import warp as wp

_MJCF = r"""
<mujoco>
  <worldbody>
    <body>
      <geom size=".1"/>
      <joint name="hinge"/>
    </body>
  </worldbody>
  <actuator>
    <motor joint="hinge"/>
  </actuator>
</mujoco>
"""

@wp.kernel
def _ctrl_callback(ctrl_out: wp.array2d(dtype=float)):
  worldid = wp.tid()
  ctrl_out[worldid, 0] = 2.0

def ctrl_callback(m, d):
  wp.launch(_ctrl_callback, dim=(d.nworld,), outputs=[d.ctrl])

mjm = mujoco.MjModel.from_xml_string(_MJCF)
m = mjw.put_model(mjm)
d = mjw.make_data(mjm)

m.callback.control = ctrl_callback
mjw.step(m, d)
assert d.ctrl.numpy()[0, 0] == 2.0

Box-box collisions#

By default, box-box collisions use the general-purpose convex collision pipeline (GJK/EPA). A specialized primitive collider based on engine_collision_box.c is available by setting the NATIVECCD disable flag:

m.opt.disableflags |= mjw.DisableBit.NATIVECCD

The specialized collider generates up to 8 contact points, compared to up to 4 for the convex pipeline, and may improve contact stability for tasks involving box stacking or manipulation.

MuJoCo Warp (MJWarp)#

Tutorial notebook#

When To Use MJWarp?#

High throughput#

Low latency#

Complex scenes#

Differentiability#

Installation#

Basic Usage#

Structs#

Batch sizes#

Functions#

Minimal example#

Command line scripts#

Feature Parity#

Performance Tuning#

Graph capture#

Batch sizes#

Solver iterations#

Contact sensor matching#

Parallel linesearch#

Memory#

Batched Model Fields#

Modifying fields#

Per-world meshes#

Batch Rendering#

Basic Usage#

Benchmarks#

Notes#

Frequently Asked Questions#

Learning frameworks#

Features#

Compilation#

Differences from MuJoCo#

Warmstart#

Inertia matrix factorization#

Options#

SDF plugins#

Physics callbacks#

Box-box collisions#

Batched `Model` Fields#