MuJoCo Accelerated: MJX & MJWarp GPU Backends
[!TIP]
Official MuJoCo documentation (doc/*.rst): Read
../../mjx.rst (.. _Mjx:, .. _MjxWarp:, .. _MjxJAX:),
../../mjx_api.rst,
../../mjwarp/index.rst (.. _MJW:), and
../../mjwarp/api.rst directly when in a MuJoCo
checkout (reserve .. _Label: grep for large files like python.rst).
Outside a checkout, the same doc/*.rst tree ships offline in
pip install mujoco at pathlib.Path(mujoco.__file__).parent / 'doc';
otherwise fetch https://mujoco.readthedocs.io/en/latest/<path>.html (or
/en/stable/ for released pages; Sphinx anchors lowercase the label, e.g.
.. _PyBatch: → #pybatch) and browse repo paths (../../../) under
https://github.com/google-deepmind/mujoco/tree/main/.
Related skills:
- mujoco-python — CPU physics simulation, kinematics,
and full solver suite.
- mujoco-parity — Cross-engine physics parity checks
between CPU MuJoCo and MJWarp/MJX (
forward isolation, contact matching,
and float32 baseline rollouts).
- mujoco-studio — Interactive simulation and Studio
viewer plugins.
[!IMPORTANT]
MJX and MJWarp are Python-only accelerated backends. Standard CPU
simulation runs in full float64 via import mujoco.
1. Backend Comparison: C++ vs MJX-JAX vs MJX-Warp vs MJWarp
| Feature |
C++ Engine (Default) |
MJX-JAX (impl='jax') |
MJX-Warp (impl='warp') |
MJWarp (CUDA) |
| Import |
import mujoco (mujoco.batch, mujoco.rollout) |
from mujoco import mjx |
from mujoco import mjx |
import mujoco_warp as mjw |
| Hardware |
CPU |
CPU / GPU / TPU |
NVIDIA GPU |
NVIDIA GPU |
| Precision |
float64 |
float32 |
float32 |
float32 |
| Differentiable |
❌ No |
✅ Yes (JAX gradients) |
❌ No |
❌ No |
| Batch Stepping |
✅ mujoco.batch.Batch & mujoco.rollout (CPU threads) |
✅ jax.vmap |
✅ jax.vmap (CUDA kernels) |
✅ Native CUDA batches |
| Solvers |
Newton, CG, PGS, noslip |
Newton, CG |
Newton, CG |
Newton, CG |
| Constraint Islands |
✅ Yes |
❌ No |
✅ Yes |
✅ Yes |
| Plugins |
All engine plugins |
❌ No |
SDF plugins only |
SDF plugins only |
See mjx.rst (.. _MjxWarp:, .. _MjxJAX:,
.. _MjxFeatureParity:) and mjwarp/index.rst
(.. _mjwDiff:) for full feature parity and solver/collision details. For
multi-threaded CPU batch simulation supporting float64, noslip/PGS, and
all engine plugins without JAX or CUDA, use stateful mujoco.batch.Batch (for
interactive RL stepping, subset resets, and per-sim domain randomization — see
python.rst, .. _PyBatch:) or stateless mujoco.rollout
(.. _PyRollout:, and programming/simulation.rst,
.. _siBatch:).
2. MJX Workflow (JAX)
MJX translates MuJoCo data structures into immutable JAX pytrees that execute
natively on GPUs and TPUs.
import mujoco
from mujoco import mjx
# 1. Load CPU model and data
model = mujoco.MjModel.from_xml_path('scene.xml')
data = mujoco.MjData(model)
# 2. Transfer to device (GPU/TPU)
mjx_model = mjx.put_model(model)
mjx_data = mjx.put_data(model, data)
# 3. Step on device (functional paradigm returns new data)
mjx_data = mjx.step(mjx_model, mjx_data)
# 4. Copy data back to host CPU
mjx.get_data_into(data, model, mjx_data)
print(data.body('torso').xpos)
On NVIDIA GPUs when JAX autodiff is not required, pass impl='warp' to
mjx.put_model(model, impl='warp') and
mjx.make_data(model, impl='warp', naconmax=..., njmax=...) (pip install mujoco-mjx[warp]) to run MJWarp's CUDA contact/constraint kernels inside JAX
(jax.jit / jax.vmap); use Data.where(...) for conditional per-world resets
and read mjx.rst (.. _MjxWarp:, .. _MjxWarpGraphModes:,
.. _MjxPerformanceWarp:) for CUDA graph modes and batch rendering.
3. High-Throughput Rollouts: jax.vmap and jax.lax.scan
Parallel Batches (jax.vmap)
import jax
from mujoco import mjx
# Vectorize step over batch dimension (model unbatched, data batched)
batched_step = jax.vmap(mjx.step, in_axes=(None, 0))
# JIT-compile for maximum device throughput
batched_step_jit = jax.jit(batched_step)
batch_data = batched_step_jit(mjx_model, batch_data)
Multi-Step Trajectory Rollouts (jax.lax.scan)
For RL environment steps, avoid Python loops by scanning across steps:
def rollout(model: mjx.Model, init_data: mjx.Data, ctrl_sequence: jax.Array):
def step_fn(d, ctrl):
d = d.replace(ctrl=ctrl)
d = mjx.step(model, d)
return d, (d.qpos, d.sensordata)
final_data, trajectory = jax.lax.scan(step_fn, init_data, ctrl_sequence)
return final_data, trajectory
4. Static Contact and Constraint Sizing
JAX requires static array shapes, so contact and constraint buffers are sized
when the device data is created, not per step.
MJX (default JAX implementation)
The contact buffer is sized from the model, not from data.ncon: every geom
pair that can collide (after contype/conaffinity, exclude and pair
filtering) gets a fixed number of contact slots, which is why large scenes
produce large efc_* arrays and slow compilation. put_data raises a
ValueError if the CPU MjData already holds more contacts than this capacity.
The njmax argument of mjx.put_data is ignored here.
To shrink the buffer, cap it with custom numerics in the MJCF:
<custom>
<!-- keep only the N closest geom pairs (bounding spheres) per geom-type pair -->
<numeric name="max_geom_pairs" data="16"/>
<!-- keep only the N deepest contacts per condim group -->
<numeric name="max_contact_points" data="32"/>
</custom>
Also reduce the number of candidate pairs with contype/conaffinity and
<exclude>, and prefer primitive geoms over meshes.
MJWarp (and mjx.put_data(..., impl='warp'))
Buffers are sized by explicit arguments; omitted values fall back to defaults
derived from the model:
nconmax: contacts per world (MJWarp put_data/make_data).
naconmax: contacts across all worlds.
njmax: constraint rows (efc) per world, a hard cap per world.
put_data raises if the input MjData exceeds a capacity. During stepping,
excess contacts/constraints are dropped, a warning is printed, and the per-world
d.overflow bitmask (mjw.OverflowType) is set, so size for the peak.
5. MJX Named Access & bind()
MJX provides declarative named access over JAX arrays via bind():
spec = mujoco.MjSpec.from_file('model.xml')
model = spec.compile()
data = mujoco.MjData(model)
mx = mjx.put_model(model)
dx = mjx.put_data(model, data)
# Read body positions from JAX device arrays
torso_pos = dx.bind(mx, spec.body('torso')).xpos
# Functional update of control inputs
dx = dx.bind(mx, spec.actuators).set('ctrl', [1.0, 0.5])
6. MJWarp Guide (NVIDIA CUDA)
MJWarp (mujoco_warp) provides a CUDA-native simulator optimized for maximum
batch throughput on NVIDIA hardware without JAX:
import mujoco
import mujoco_warp as mjw
model = mujoco.MjModel.from_xml_path('scene.xml')
data = mujoco.MjData(model)
# Transfer to CUDA: one model, nworld copies of the data
mjw_model = mjw.put_model(model)
mjw_data = mjw.put_data(model, data, nworld=4096, nconmax=64, njmax=256)
# or allocate fresh: mjw.make_data(model, nworld=4096, nconmax=64, njmax=256)
# Advance all 4096 environments in parallel (in place)
mjw.step(mjw_model, mjw_data)
# Fields are Warp arrays with a leading world dimension
qpos = mjw_data.qpos.numpy() # (4096, nq)
# Copy one world back to CPU MjData
mjw.get_data_into(data, model, mjw_data, world_id=0)
7. Feature Parity, Tuning & Gotchas
Instead of relying on static feature snapshots, check the authoritative parity
and performance sections in the RST manual when choosing colliders, solvers, or
tuning throughput:
- Supported geometries, solvers & differences from CPU MuJoCo:
mjx.rst (
.. _MjxFeatureParity:, .. _MjxSharpBits:) and
mjwarp/index.rst (.. _mjwDiff:,
.. _mjwPairwise:, .. _mjwBoxBox:). Notably, noslip and PGS solvers
are CPU-only, and MJX-JAX solves constraints monolithically without
constraint islands.
- Performance tuning, CUDA graphs & domain randomization:
mjx.rst (
.. _MjxPerformance:, .. _MjxPerformanceJAX:,
.. _MjxPerformanceWarp:), mjwarp/index.rst
(.. _mjwPerf:, .. _mjwGC:, .. _mjwOverflow:, .. _mjwBatch:), and
modeling.rst (.. _CTuning:, .. _CTuningGlobal:,
.. _CTuningModel:) for general timestep, solver/iteration, friction-cone,
and primitive-vs-mesh collision trade-offs.
High-Frequency Gotchas
- Float32 precision: GPU backends run in single precision (
float32).
Stiff systems may experience numerical drift or NaN errors. Counteract this
by increasing model.opt.iterations or decreasing model.opt.timestep.
For comparing float32 GPU physics against CPU MuJoCo, refer to
mujoco-parity.
- Immutability in MJX:
mjx.Data is a JAX pytree. In-place assignments
fail; use .replace():mjx_data = mjx_data.replace(qpos=new_qpos)
- Compilation caching: Avoid long recompilations on startup by enabling
persistent XLA caching:
import jax
jax.config.update('jax_compilation_cache_dir', '/tmp/jax_cache')
8. Key References
Documentation
- MJX Guide (
MJX-JAX & MJX-Warp) (.. _Mjx:,
.. _MjxWarp:, .. _MjxJAX:, .. _MjxFeatureParity:,
.. _MjxPerformance:, .. _MjxSharpBits:) and
MJX API Reference
- MuJoCo Warp (MJWarp) Guide (
.. _MJW:,
.. _MJW_Usage:, .. _mjwPerf:, .. _mjwOverflow:, .. _mjwBatch:,
.. _mjwDiff:) and MJWarp API Reference
- Python Bindings (
mujoco.batch & mujoco.rollout)
(.. _PyNamed:, .. _PyBatch:, .. _PyRollout:),
Model Tuning Guide (.. _CTuning:), and
MuJoCo Menagerie (.. _Menagerie:)
Notebooks, Models & RL Ecosystem
1---2name: mujoco-accelerated3description: >-4 MuJoCo GPU and TPU batch simulation with MJX (mujoco.mjx in JAX) and MJWarp5 (mujoco_warp in CUDA): parallel rollouts6 (jax.vmap), trajectory scan (jax.lax.scan), differentiability, device memory7 (put_model, put_data), static contact sizing (max_contact_points,8 max_geom_pairs, nconmax, njmax), float32 precision, XLA compilation caching.9 Use only for MuJoCo MJX or MJWarp GPU/TPU batch rollouts, RL environments, and10 differentiable MuJoCo physics. Don't use for custom non-MuJoCo JAX or CUDA11 Warp simulations (e.g. custom 2D games, molecular dynamics, orbital mechanics,12 DEM granular solvers, or shallow-water PDEs that do not import `mujoco.mjx` or13 `mujoco_warp`), CPU simulation (mujoco-python), cross-engine parity debugging14 (mujoco-parity), rendering (mujoco-rendering), or GUI (mujoco-gui).15metadata:16 source: https://skillmd.com/skills/google-deepmind/mujoco-accelerated17---1819# MuJoCo Accelerated: MJX & MJWarp GPU Backends2021> [!TIP]22>23> **Official MuJoCo documentation (`doc/*.rst`):** Read24> [`../../mjx.rst`](../../mjx.rst) (`.. _Mjx:`, `.. _MjxWarp:`, `.. _MjxJAX:`),25> [`../../mjx_api.rst`](../../mjx_api.rst),26> [`../../mjwarp/index.rst`](../../mjwarp/index.rst) (`.. _MJW:`), and27> [`../../mjwarp/api.rst`](../../mjwarp/api.rst) directly when in a MuJoCo28> checkout (reserve `.. _Label:` grep for large files like `python.rst`).29> Outside a checkout, the same `doc/*.rst` tree ships offline in30> `pip install mujoco` at `pathlib.Path(mujoco.__file__).parent / 'doc'`;31> otherwise fetch `https://mujoco.readthedocs.io/en/latest/<path>.html` (or32> `/en/stable/` for released pages; Sphinx anchors lowercase the label, e.g.33> `.. _PyBatch:` → `#pybatch`) and browse repo paths (`../../../`) under34> `https://github.com/google-deepmind/mujoco/tree/main/`.35>36> **Related skills:**37>38> - [mujoco-python](../python/SKILL.md) — CPU physics simulation, kinematics,39> and full solver suite.40> - [mujoco-parity](../parity/SKILL.md) — Cross-engine physics parity checks41> between CPU MuJoCo and MJWarp/MJX (`forward` isolation, contact matching,42> and `float32` baseline rollouts).43> - [mujoco-studio](../studio/SKILL.md) — Interactive simulation and Studio44> viewer plugins.4546> [!IMPORTANT]47>48> MJX and MJWarp are **Python-only** accelerated backends. Standard CPU49> simulation runs in full `float64` via `import mujoco`.5051## 1. Backend Comparison: C++ vs MJX-JAX vs MJX-Warp vs MJWarp5253| Feature | C++ Engine (Default) | MJX-JAX (`impl='jax'`) | MJX-Warp (`impl='warp'`) | MJWarp (CUDA) |54| :--- | :--- | :--- | :--- | :--- |55| **Import** | `import mujoco` (`mujoco.batch`, `mujoco.rollout`) | `from mujoco import mjx` | `from mujoco import mjx` | `import mujoco_warp as mjw` |56| **Hardware** | CPU | CPU / GPU / TPU | NVIDIA GPU | NVIDIA GPU |57| **Precision** | float64 | float32 | float32 | float32 |58| **Differentiable** | ❌ No | ✅ Yes (JAX gradients) | ❌ No | ❌ No |59| **Batch Stepping** | ✅ `mujoco.batch.Batch` & `mujoco.rollout` (CPU threads) | ✅ `jax.vmap` | ✅ `jax.vmap` (CUDA kernels) | ✅ Native CUDA batches |60| **Solvers** | Newton, CG, PGS, **noslip** | Newton, CG | Newton, CG | Newton, CG |61| **Constraint Islands**| ✅ Yes | ❌ No | ✅ Yes | ✅ Yes |62| **Plugins** | All engine plugins | ❌ No | SDF plugins only | SDF plugins only |6364See [mjx.rst](../../mjx.rst) (`.. _MjxWarp:`, `.. _MjxJAX:`,65`.. _MjxFeatureParity:`) and [mjwarp/index.rst](../../mjwarp/index.rst)66(`.. _mjwDiff:`) for full feature parity and solver/collision details. For67multi-threaded CPU batch simulation supporting `float64`, `noslip`/`PGS`, and68all engine plugins without JAX or CUDA, use stateful `mujoco.batch.Batch` (for69interactive RL stepping, subset resets, and per-sim domain randomization — see70[python.rst](../../python.rst), `.. _PyBatch:`) or stateless `mujoco.rollout`71(`.. _PyRollout:`, and [programming/simulation.rst](../../programming/simulation.rst),72`.. _siBatch:`).7374## 2. MJX Workflow (JAX)7576MJX translates MuJoCo data structures into immutable JAX pytrees that execute77natively on GPUs and TPUs.7879```python80import mujoco81from mujoco import mjx8283# 1. Load CPU model and data84model = mujoco.MjModel.from_xml_path('scene.xml')85data = mujoco.MjData(model)8687# 2. Transfer to device (GPU/TPU)88mjx_model = mjx.put_model(model)89mjx_data = mjx.put_data(model, data)9091# 3. Step on device (functional paradigm returns new data)92mjx_data = mjx.step(mjx_model, mjx_data)9394# 4. Copy data back to host CPU95mjx.get_data_into(data, model, mjx_data)96print(data.body('torso').xpos)97```9899On NVIDIA GPUs when JAX autodiff is not required, pass `impl='warp'` to100`mjx.put_model(model, impl='warp')` and101`mjx.make_data(model, impl='warp', naconmax=..., njmax=...)` (`pip install102mujoco-mjx[warp]`) to run MJWarp's CUDA contact/constraint kernels inside JAX103(`jax.jit` / `jax.vmap`); use `Data.where(...)` for conditional per-world resets104and read [mjx.rst](../../mjx.rst) (`.. _MjxWarp:`, `.. _MjxWarpGraphModes:`,105`.. _MjxPerformanceWarp:`) for CUDA graph modes and batch rendering.106107## 3. High-Throughput Rollouts: jax.vmap and jax.lax.scan108109### Parallel Batches (`jax.vmap`)110111```python112import jax113from mujoco import mjx114115# Vectorize step over batch dimension (model unbatched, data batched)116batched_step = jax.vmap(mjx.step, in_axes=(None, 0))117118# JIT-compile for maximum device throughput119batched_step_jit = jax.jit(batched_step)120batch_data = batched_step_jit(mjx_model, batch_data)121```122123### Multi-Step Trajectory Rollouts (`jax.lax.scan`)124125For RL environment steps, avoid Python loops by scanning across steps:126127```python128def rollout(model: mjx.Model, init_data: mjx.Data, ctrl_sequence: jax.Array):129 def step_fn(d, ctrl):130 d = d.replace(ctrl=ctrl)131 d = mjx.step(model, d)132 return d, (d.qpos, d.sensordata)133134 final_data, trajectory = jax.lax.scan(step_fn, init_data, ctrl_sequence)135 return final_data, trajectory136```137138## 4. Static Contact and Constraint Sizing139140JAX requires static array shapes, so contact and constraint buffers are sized141when the device data is created, not per step.142143### MJX (default JAX implementation)144145The contact buffer is sized **from the model**, not from `data.ncon`: every geom146pair that can collide (after `contype`/`conaffinity`, `exclude` and `pair`147filtering) gets a fixed number of contact slots, which is why large scenes148produce large `efc_*` arrays and slow compilation. `put_data` raises a149`ValueError` if the CPU `MjData` already holds more contacts than this capacity.150The `njmax` argument of `mjx.put_data` is ignored here.151152To shrink the buffer, cap it with custom numerics in the MJCF:153154```xml155<custom>156 <!-- keep only the N closest geom pairs (bounding spheres) per geom-type pair -->157 <numeric name="max_geom_pairs" data="16"/>158 <!-- keep only the N deepest contacts per condim group -->159 <numeric name="max_contact_points" data="32"/>160</custom>161```162163Also reduce the number of candidate pairs with `contype`/`conaffinity` and164`<exclude>`, and prefer primitive geoms over meshes.165166### MJWarp (and `mjx.put_data(..., impl='warp')`)167168Buffers are sized by explicit arguments; omitted values fall back to defaults169derived from the model:170171- `nconmax`: contacts per world (MJWarp `put_data`/`make_data`).172- `naconmax`: contacts across all worlds.173- `njmax`: constraint rows (`efc`) **per world**, a hard cap per world.174175`put_data` raises if the input `MjData` exceeds a capacity. During stepping,176excess contacts/constraints are dropped, a warning is printed, and the per-world177`d.overflow` bitmask (`mjw.OverflowType`) is set, so size for the peak.178179## 5. MJX Named Access & `bind()`180181MJX provides declarative named access over JAX arrays via `bind()`:182183```python184spec = mujoco.MjSpec.from_file('model.xml')185model = spec.compile()186data = mujoco.MjData(model)187188mx = mjx.put_model(model)189dx = mjx.put_data(model, data)190191# Read body positions from JAX device arrays192torso_pos = dx.bind(mx, spec.body('torso')).xpos193194# Functional update of control inputs195dx = dx.bind(mx, spec.actuators).set('ctrl', [1.0, 0.5])196```197198## 6. MJWarp Guide (NVIDIA CUDA)199200MJWarp (`mujoco_warp`) provides a CUDA-native simulator optimized for maximum201batch throughput on NVIDIA hardware without JAX:202203```python204import mujoco205import mujoco_warp as mjw206207model = mujoco.MjModel.from_xml_path('scene.xml')208data = mujoco.MjData(model)209210# Transfer to CUDA: one model, nworld copies of the data211mjw_model = mjw.put_model(model)212mjw_data = mjw.put_data(model, data, nworld=4096, nconmax=64, njmax=256)213# or allocate fresh: mjw.make_data(model, nworld=4096, nconmax=64, njmax=256)214215# Advance all 4096 environments in parallel (in place)216mjw.step(mjw_model, mjw_data)217218# Fields are Warp arrays with a leading world dimension219qpos = mjw_data.qpos.numpy() # (4096, nq)220221# Copy one world back to CPU MjData222mjw.get_data_into(data, model, mjw_data, world_id=0)223```224225## 7. Feature Parity, Tuning & Gotchas226227Instead of relying on static feature snapshots, check the authoritative parity228and performance sections in the RST manual when choosing colliders, solvers, or229tuning throughput:230231- **Supported geometries, solvers & differences from CPU MuJoCo**:232 [mjx.rst](../../mjx.rst) (`.. _MjxFeatureParity:`, `.. _MjxSharpBits:`) and233 [mjwarp/index.rst](../../mjwarp/index.rst) (`.. _mjwDiff:`,234 `.. _mjwPairwise:`, `.. _mjwBoxBox:`). Notably, `noslip` and `PGS` solvers235 are CPU-only, and MJX-JAX solves constraints monolithically without236 constraint islands.237- **Performance tuning, CUDA graphs & domain randomization**:238 [mjx.rst](../../mjx.rst) (`.. _MjxPerformance:`, `.. _MjxPerformanceJAX:`,239 `.. _MjxPerformanceWarp:`), [mjwarp/index.rst](../../mjwarp/index.rst)240 (`.. _mjwPerf:`, `.. _mjwGC:`, `.. _mjwOverflow:`, `.. _mjwBatch:`), and241 [modeling.rst](../../modeling.rst) (`.. _CTuning:`, `.. _CTuningGlobal:`,242 `.. _CTuningModel:`) for general timestep, solver/iteration, friction-cone,243 and primitive-vs-mesh collision trade-offs.244245### High-Frequency Gotchas2462471. **Float32 precision**: GPU backends run in single precision (`float32`).248 Stiff systems may experience numerical drift or NaN errors. Counteract this249 by increasing `model.opt.iterations` or decreasing `model.opt.timestep`.250 For comparing `float32` GPU physics against CPU MuJoCo, refer to251 [`mujoco-parity`](../parity/SKILL.md).2522. **Immutability in MJX**: `mjx.Data` is a JAX pytree. In-place assignments253 fail; use `.replace()`:254 ```python255 mjx_data = mjx_data.replace(qpos=new_qpos)256 ```2573. **Compilation caching**: Avoid long recompilations on startup by enabling258 persistent XLA caching:259 ```python260 import jax261 jax.config.update('jax_compilation_cache_dir', '/tmp/jax_cache')262 ```263264## 8. Key References265266### Documentation267268- [MJX Guide (`MJX-JAX` & `MJX-Warp`)](../../mjx.rst) (`.. _Mjx:`,269 `.. _MjxWarp:`, `.. _MjxJAX:`, `.. _MjxFeatureParity:`,270 `.. _MjxPerformance:`, `.. _MjxSharpBits:`) and271 [MJX API Reference](../../mjx_api.rst)272- [MuJoCo Warp (MJWarp) Guide](../../mjwarp/index.rst) (`.. _MJW:`,273 `.. _MJW_Usage:`, `.. _mjwPerf:`, `.. _mjwOverflow:`, `.. _mjwBatch:`,274 `.. _mjwDiff:`) and [MJWarp API Reference](../../mjwarp/api.rst)275- [Python Bindings (`mujoco.batch` & `mujoco.rollout`)](../../python.rst)276 (`.. _PyNamed:`, `.. _PyBatch:`, `.. _PyRollout:`),277 [Model Tuning Guide](../../modeling.rst) (`.. _CTuning:`), and278 [MuJoCo Menagerie](../../models.rst) (`.. _Menagerie:`)279280### Notebooks, Models & RL Ecosystem281282- **Tutorials & tests**: `mjx/tutorial.ipynb` (`.. _MjxNotebook:` in283 [mjx.rst](../../mjx.rst)),284 [MJX Support Tests](../../../mjx/mujoco/mjx/_src/support_test.py),285 [Humanoid Models](../../../model/humanoid/humanoid.xml), and286 [Benchmark Test Models](../../../test/benchmark/testdata/100_humanoids_chol.xml)287- **RL & simulation ecosystem** (see288 [mjwarp/index.rst](../../mjwarp/index.rst), `.. _mjwFAQ:`):289 - [MuJoCo Playground (`google-deepmind/mujoco_playground`)](https://github.com/google-deepmind/mujoco_playground)290 — JAX/MJX (`MJX-JAX` & `MJX-Warp`) RL environments for locomotion,291 manipulation, dexterous hands, and vision.292 - [`mjlab` (`mujocolab/mjlab`)](https://github.com/mujocolab/mjlab) —293 Isaac Lab-style manager API for training policies directly with294 `mujoco_warp` and PyTorch.295 - [MuJoCo Warp (`google-deepmind/mujoco_warp`)](https://github.com/google-deepmind/mujoco_warp)296 — Standalone `mujoco_warp` repository (`notebooks/tutorial.ipynb`,297 benchmarks, and Warp kernels).298 - [MuJoCo MPC (`google-deepmind/mujoco_mpc`)](https://github.com/google-deepmind/mujoco_mpc)299 — Real-time model-predictive control and MJX planning baselines.