2.3.3

General

1. Improvements to implicit integration:

   • The derivatives of the RNE algorithm are now computed using sparse math, leading to significant speed improvements for large models when using the implicit integrator.
   • A new integrator called implicitfast was added. It is similar to the existing implicit integrator, but skips the derivatives of Coriolis and centripetal forces. See the numerical integration section for a detailed motivation and discussion. The implicitfast integrator is recommended for all new models and will become the default integrator in a future version.

   The table below shows the compute cost of the 627-DoF humanoid100 model using different integrators. "implicit (old)" uses dense RNE derivatives, "implicit (new)" is after the sparsification mentioned above. Timings were measured on a single core of an AMD 3995WX CPU.

timing	Euler	implicitfast	implicit (new)	implicit (old)
one step (ms)	0.5	0.53	0.77	5.0
steps/second	2000	1900	1300	200

2. Added a collision mid-phase for pruning geoms in body pairs, see documentation for more details.
   This is based on static AABB bounding volume hierarchy (a BVH binary tree) in the body inertial frame. The GIF on the right is cut from this longer video.
3. The mjd_transitionFD function no longer triggers sensor calculation unless explicitly requested.
4. Corrected the spelling of the inteval attribute to interval in the mjLROpt struct.
5. Mesh texture and normal mappings are now 3-per-triangle rather than 1-per-vertex. Mesh vertices are no longer duplicated in order to circumvent this limitation as they previously were.
6. The non-zeros for the sparse constraint Jacobian matrix are now precounted and used for matrix memory allocation.
   For instance, the constraint Jacobian matrix from the humanoid100 model, which previously required ~500,000 mjtNums, now only requires ~6000. Very large models can now load and run with the CG solver.
7. Modified mju_error and mju_warning to be variadic functions (support for printf-like arguments). The functions mju_error_i, mju_error_s, mju_warning_i, and mju_warning_s are now deprecated.
8. Implemented a performant mju_sqrMatTDSparse function that doesn't require dense memory allocation.
9. Added mj_stackAllocInt to get correct size for allocating ints on mjData stack. Reducing stack memory usage by 10% - 15%.

Python bindings

10. Fixed IPython history corruption when using viewer.launch_repl. The launch_repl function now provides seamless continuation of an IPython interactive shell session, and is no longer considered experimental feature.
11. Added viewer.launch_passive which launches the interactive viewer in a passive, non-blocking mode. Calls to launch_passive return immediately, allowing user code to continue execution, with the viewer automatically reflecting any changes to the physics state. (Note that this functionality is currently in experimental/beta stage, and is not yet described in our viewer documentation.)
12. Added the mjpython launcher for macOS, which is required for viewer.launch_passive to function there.
13. Removed efc_ fields from joint indexers. Since the introduction of arena memory, these fields now have dynamic sizes that change between time steps depending on the number of active constraints, breaking strict correspondence between joints and efc_ rows.
14. Added a number of missing fields to the bindings of mjVisual and mjvPerturb structs.

Simulate

15. Implemented a workaround for broken VSync on macOS so that the frame rate is correctly capped when the Vertical Sync toggle is enabled.

16. Added optional labels to contact visualization, indicating which two geoms are contacting (names if defined, ids otherwise). This can be useful in cluttered scenes.

2.3.2

General

1. A more performant mju_transposeSparse has been implemented that doesn't require dense memory allocation. For a constraint Jacobian matrix from the humanoid100.xml model, this function is 35% faster.
2. The function mj_name2id is now implemented using a hash function instead of a linear search for better performance.
3. Geom names are now parsed from URDF. Any duplicate names are ignored. mj_printData output now contains contacting geom names.

Bug fixes

4. Fixed a bug that for shellinertia equal to true caused the mesh orientation to be overwritten by the principal components of the shell inertia, while the vertex coordinates are rotated using the volumetric inertia. Now the volumetric inertia orientation is used also in the shell case.
5. Fixed misalignment bug in mesh-to-primitive fitting when using the bounding box fitting option fitaabb.
6. The launch_repl functionality in the Python viewer has been fixed.
7. Set time correctly in mjd_transitionFD, to support time-dependent user code.
8. Fixed sensor data dimension validation when user type sensors are present.
9. Fixed incorrect plugin error message when a null nsensordata callback is encountered during model compilation.
10. Correctly end the timer (TM_END) mj_fwdConstraint returns early.
11. Fixed an infinite loop in mj_deleteFileVFS.

Simulate

12. Increased precision of simulate sensor plot y-axis by 1 digit #719.
13. Body labels are now drawn at the body frame rather than inertial frame, unless inertia is being visualised.

Plugins

14. The reset callback now receives instance-specific plugin_state and plugin_data as arguments, rather than the entire mjData. Since reset is called inside mj_resetData before any physics forwarding call has been made, it is an error to read anything from mjData at this stage.
15. The capabilities field in mjpPlugin is renamed capabilityflags to more clearly indicate that this is a bit field.

2.3.1

Python bindings

1. The simulate GUI is now available through the mujoco Python package as mujoco.viewer. See documentation for details. (Contribution by Levi Burner.)
2. The Renderer class from the MuJoCo tutorial Colab is now available directly in the native Python bindings.

General

3. The tendon springlength attribute can now take two values. Given two non-decreasing values, springlength specifies a deadband range for spring stiffness. If the tendon length is between the two values, the force is 0. If length is outside this range, the force behaves like a regular spring, with the spring resting length corresponding to the nearest springlength value. This can be used to create tendons whose limits are enforced by springs rather than constraints, which are cheaper and easier to analyse. See tendon_springlength.xml example model. This is a minor breaking API change. mjModel.tendon_lengthspring now has size ntendon x 2 rather than ntendon x 1.

4. Removed the requirement that stateless actuators come before stateful actuators.

5. Added mju_fill, mju_symmetrize and mju_eye utility functions.

6. Added gravcomp attribute to body implementing gravity compensation and buoyancy. See balloons.xml example model.

7. Renamed the cable plugin library to elasticity.

8. Added actdim attribute to general actuators. Values greater than 1 are only allowed for user, as native activation dynamics are all scalar. Added example test implementing 2nd-order activation dynamics to engine_forward_test.cc.

9. Improved particle composite type, which now permits a user-specified geometry and multiple joints. See the two new examples: particle_free.xml and particle_free2d.xml.

10. Performance improvements for non-AVX configurations:

    • 14% faster mj_solveLD using restrict. See engine_core_smooth_benchmark_test.
    • 50% faster mju_dotSparse using manual loop unroll. See engine_util_sparse_benchmark_test.

11. Added new solid passive force plugin.

    • This is new force field compatible with the composite particles.
    • Generates a tetrahedral mesh having particles with mass concentrated at vertices.
    • Uses a piecewise-constant strain model equivalent to finite elements but expressed in a coordinate-free
      formulation. This implies that all quantities can be precomputed except edge elongation, as in a mass-spring model.
    • Only suitable for small strains (large displacements but small deformations). Tetrahedra may invert if subject to
      large loads.

12. Added API functions mj_loadPluginLibrary and mj_loadAllPluginLibraries. The first function is identical to dlopen on a POSIX system, and to LoadLibraryA on Windows. The second function scans a specified directory for all dynamic libraries file and loads each library found. Dynamic libraries opened by these functions are assumed to register one or more MuJoCo plugins on load.

13. Added an optional visualize callback to plugins, which is called during mjv_updateScene. This callback allows custom plugin visualizations. Enable stress visualization for the Cable plugin as an example.

14. Sensors of type user no longer require objtype, objname and needstage. If unspecified, the objtype is now mjOBJ_UNKNOWN. user sensors datatype default is now real, needstage default is now acc.

15. Added support for capsules in URDF import.

16. On macOS, issue an informative error message when run under Rosetta 2 translation on an Apple Silicon machine. Pre-built MuJoCo binaries make use of AVX instructions on x86-64 machines, which is not supported by Rosetta 2. (Before this version, users only get a cryptic "Illegal instruction" message.)

Bug fixes

17. Fixed bug in mj_addFileVFS that was causing the file path to be ignored (introduced in 2.1.4).

Simulate

18. Renamed the directory in which the simulate application searches for plugins from plugin to mujoco_plugin.
19. Mouse force perturbations are now applied at the selection point rather than the body center of mass.

---

Python bindings 2.3.1.post1 (6 December 2022)

• Fix type annotations in mujoco.viewer that were causing errors in Python 3.7 and 3.8.

2.3.0

General

1. The contact array and arrays prefixed with efc_ in mjData were moved out of the buffer into a new arena memory space. These arrays are no longer allocated with fixed sizes when mjData is created. Instead, the exact memory requirement is determined during each call to mj_forward (specifically, in mj_collision and mj_makeConstraint) and the arrays are allocated from the arena space. The stack now also shares its available memory with arena. This change reduces the memory footprint of mjData in models that do not use the PGS solver, and will allow for significant memory reductions in the future. See the Memory allocation section for details.

2. Added colab notebook tutorial showing how to balance the humanoid on one leg with a Linear Quadratic Regulator. The
   notebook uses MuJoCo's native Python bindings, and includes a draft Renderer class, for easy rendering in Python. Try it yourself:

3. Updates to humanoid model:

   • Added two keyframes (stand-on-one-leg and squat).
   • Increased maximum hip flexion angle.
   • Added hamstring tendons which couple the hip and knee at high hip flexion angles.
   • General cosmetic improvements, including improved use of defaults and better naming scheme.

4. Added mju_boxQP and allocation function mju_boxQPmalloc for solving the box-constrained Quadratic Program: $x^* = \text{argmin} ; \tfrac{1}{2} x^T H x + x^T g \quad \text{s.t.} \quad l \le x \le u$ The algorithm, introduced in Tassa et al. 2014, converges after 2-5 Cholesky factorisations, independent of problem size.

5. Added mju_mulVecMatVec to multiply a square matrix $M$ with vectors $x$ and $y$ on both sides. The function returns $x^TMy$.

6. Added new plugin API. Plugins allow developers to extend MuJoCo's capability without modifying core engine code. The plugin mechanism is intended to replace the existing callbacks, though these will remain for the time being as an option for simple use cases and backward compatibility. The new mechanism manages stateful plugins and supports multiple plugins from different sources, allowing MuJoCo extensions to be introduced in a modular fashion, rather than as global overrides. Note the new mechanism is currently undocumented except in code, as we test it internally. If you are interested in using the plugin mechanism, please get in touch first.

7. Added assetdir compiler option, which sets the values of both meshdir and texturedir. Values in the latter attributes take precedence over assetdir.

8. Added realtime option to visual for starting a simulation at a slower speed.

9. Added new cable composite type:

   • Cable elements are connected with ball joints.
   • The initial parameter specifies the joint at the starting boundary: free, ball, or none.
   • The boundary bodies are exposed with the names B_left and B_right.
   • The vertex initial positions can be specified directly in the XML with the parameter vertex.
   • The orientation of the body frame is the orientation of the material frame of the curve.

10. Added new cable passive force plugin:

    • Twist and bending stiffness can be set separately with the parameters twist and bend.
    • The stress-free configuration can be set to be the initial one or flat with the flag flat.
    • New cable.xml example showing the formation of plectoneme.
    • New coil.xml example showing a curved equilibrium configuration.
    • New belt.xml example showing interaction between twist and anisotropy.
    • Added test using cantilever exact solution.

Python bindings

11. Added id and name properties to named accessor objects. These provide more Pythonic API access to mj_name2id and mj_id2name respectively.

12. The length of MjData.contact is now ncon rather than nconmax, allowing it to be straightforwardly used as
    an iterator without needing to check ncon.

13. Fix a memory leak when a Python callable is installed as callback (#527).

2.2.2

General

1. Added adhesion actuators mimicking vacuum grippers and adhesive biomechanical appendages.

2. Added related example model and video.

3. Added mj_jacSubtreeCom for computing the translational Jacobian of the center-of-mass of a subtree.

4. Added torquescale and anchor attributes to weld constraints. torquescale sets the torque-to-force ratio exerted by the constraint, anchor sets the point at which the weld wrench is applied. See weld for more details.

5. Increased mjNEQDATA, the row length of equality constraint parameters in mjModel.eq_data, from 7 to 11.

6. Added visualisation of anchor points for both connect and weld constraints (activated by the 'N' key in simulate).

7. Added weld.xml showing different uses of new weld attributes.

8. Cartesian 6D end-effector control is now possible by adding a reference site to actuators with site transmission. See description of new refsite attribute in the actuator documentation and refsite.xml example model. Video

9. Added autolimits compiler option. If true, joint and tendon limited attributes and actuator ctrllimited, forcelimited and actlimited attributes will automatically be set to true if the corresponding range is defined and false otherwise.

   If autolimits="false" (the default) models where a range attribute is specified without the limited attribute will fail to compile. A future release will change the default of autolimits to true, and this compilation error allows users to catch this future change of behavior.

   This is a breaking change. In models where a range was defined but limited was unspecified, explicitly set limited to false or remove the range to maintain the current behavior of your model.

10. Added moment of inertia computation for all well-formed meshes. This option is activated by setting the compiler flag exactmeshinertia to true (defaults to false). This default may change in the future.

11. Added parameter shellinertia to geom, for locating the inferred inertia on the boundary (shell). Currently only meshes are supported.

12. For meshes from which volumetric inertia is inferred, raise error if the orientation of mesh faces is not consistent. If this occurs, fix the mesh in e.g., MeshLab or Blender.

13. Added catenary visualisation for hanging tendons. The model seen in the video can be found here.

14. Added azimuth and elevation attributes to visual/global, defining the initial orientation of the free camera at model load time.

15. Added mjv_defaultFreeCamera which sets the default free camera, respecting the above attributes.

16. simulate now supports taking a screenshot via a button in the File section or via Ctrl-P.

17. Improvements to time synchronisation in simulate, in particular report actual real-time factor if different from requested factor (if e.g., the timestep is so small that simulation cannot keep up with real-time).

18. Added a disable flag for sensors.

19. mju_mulQuat and mju_mulQuatAxis support in place computation. For example mju_mulQuat(a, a, b); sets the quaternion a equal to the product of a and b.

20. Added sensor matrices to mjd_transitionFD (note this is an API change).

Deleted/deprecated features

21. Removed distance constraints.

Bug fixes

22. Fixed rendering of some transparent geoms in reflection.
23. Fixed intvelocity defaults parsing.

2.2.1

General

1. Added mjd_transitionFD to compute efficient finite difference approximations of the state-transition and control-transition matrices, see here for more details.
2. Added derivatives for the ellipsoid fluid model.
3. Added ctrl attribute to keyframes.
4. Added clock sensor which measures time.
5. Added visualisation groups to skins.
6. Added actuator visualisation for free and ball joints and for actuators with site transmission.
7. Added visualisation for actuator activations.
8. Added <intvelocity> actuator shortcut for "integrated velocity" actuators, documented here.
9. Added <damper> actuator shortcut for active-damping actuators, documented here.
10. mju_rotVecMat and mju_rotVecMatT now support in-place multiplication.
11. mjData.ctrl values are no longer clamped in-place, remain untouched by the engine.
12. Arrays in mjData's buffer now align to 64-byte boundaries rather than 8-byte.
13. Add memory poisoning when building with Address Sanitizer (ASAN) and Memory Sanitizer (MSAN). This allows ASAN to detect reads and writes to regions in mjModel.buffer and mjData.buffer that do not lie within an array, and for MSAN to detect reads from uninitialised fields in mjData following mj_resetData.
14. Add a slider-crank example to model/.

Bug fixes

15. Activation clamping was not being applied in the implicit integrator.
16. Stricter parsing of orientation specifiers. Before this change, a specification that included both quat and an alternative specifier e.g., <geom ... quat=".1 .2 .3 .4" euler="10 20 30">, would lead to the quat being ignored and only euler being used. After this change a parse error will be thrown.
17. Stricter parsing of XML attributes. Before this change an erroneous XML snippet like <geom size="1/2 3 4"> would have been parsed as size="1 0 0" and no error would have been thrown. Now throws an error.
18. Trying to load a NaN via XML like <geom size="1 NaN 4">, while allowed for debugging purposes, will now print a warning.
19. Fixed null pointer dereference in mj_loadModel.
20. Fixed memory leaks when loading an invalid model from MJB.
21. Integer overflows are now avoided when computing mjModel buffer sizes.
22. Added missing warning string for mjWARN_BADCTRL.

Packaging

23. Changed MacOS packaging so that the copy of mujoco.framework embedded in MuJoCo.app can be used to build applications externally.

2.2.0

Open Sourcing

1. MuJoCo is now fully open-source software. Newly available top level directories are:

   1. src/: All source files. Subdirectories correspond to the modules described in the Programming chapter introduction:
      • src/engine/: Core engine.
      • src/xml/: XML parser.
      • src/user/: Model compiler.
      • src/visualize/: Abstract visualizer.
      • src/ui/: UI framework.
   2. test/: Tests and corresponding asset files.
   3. dist/: Files related to packaging and binary distribution.

2. Added contributor's guide and style guide.

General

3. Added analytic derivatives of smooth (unconstrained) dynamics forces, with respect to velocities:

   • Centripetal and Coriolis forces computed by the Recursive Newton-Euler algorithm.
   • Damping and fluid-drag passive forces.
   • Actuation forces.

4. Added implicit integrator. Using the analytic derivatives above, a new implicit-in-velocity integrator was added. This integrator lies between the Euler and Runge Kutta integrators in terms of both stability and computational cost. It is most useful for models which use fluid drag (e.g. for flying or swimming) and for models which use velocity actuators. For more details, see the Numerical Integration section.

5. Added actlimited and actrange attributes to general actuators, for clamping actuator internal states (activations). This clamping is useful for integrated-velocity actuators, see the Activation clamping section for details.

6. mjData fields qfrc_unc (unconstrained forces) and qacc_unc (unconstrained accelerations) were renamed qfrc_smooth and qacc_smooth, respectively. While "unconstrained" is precise, "smooth" is more intelligible than "unc".

7. Public headers have been moved from /include to /include/mujoco/, in line with the directory layout common in other open source projects. Developers are encouraged to include MuJoCo public headers in their own codebase via #include <mujoco/filename.h>.

8. The default shadow resolution specified by the shadowsize attribute was increased from 1024 to 4096.

9. Saved XMLs now use 2-space indents.

Bug fixes

10. Antialiasing was disabled for segmentation rendering. Before this change, if the offsamples attribute was greater than 0 (the default value is 4), pixels that overlapped with multiple geoms would receive averaged segmentation IDs, leading to incorrect or non-existent IDs. After this change offsamples is ignored during segmentation rendering.

11. The value of the enable flag for the experimental multiCCD feature was made sequential with other enable flags. Sequentiality is assumed in the simulate UI and elsewhere.

12. Fix issue of duplicated meshes when saving models with OBJ meshes using mj_saveLastXML.

---

Update 1 (27 May 2022): Replaced Linux tarball to fix RPATH on the sample binaries. No change in functionality.

2.1.5

General

1. Added an experimental feature: multi-contact convex collision detection, activated by an enable flag. See full
   description here.

Bug fixes

2. GLAD initialization logic on Linux now calls dlopen to load a GL platform dynamic library if a *GetProcAddress function is not already present in the process' global symbol table. In particular, processes that use GLFW to set up a rendering context that are not explicitly linked against libGLX.so (this applies to the Python interpreter, for example) will now work correctly rather than fail with a gladLoadGL error when mjr_makeContext is called. Fixes google-deepmind/dm_control#283.

3. In the Python bindings, named indexers for scalar fields (e.g. the ctrl field for actuators) now return a NumPy array of shape (1,) rather than (). This allows values to be assigned to these fields more straightforwardly. Fixes #238.

2.1.4

General

1. MuJoCo now uses GLAD to manage OpenGL API access instead of GLEW. On Linux, there is no longer a need to link against different GL wrangling libraries depending on whether GLX, EGL, or OSMesa is being used. Instead, users can simply use GLX, EGL, or OSMesa to create a GL context and mjr_makeContext will detect which one is being used.

2. Add visualisation for contact frames. This is useful when writing or modifying collision functions, when the actual direction of the x and y axes of a contact can be important.

Binary build

3. The _nogl dynamic library is no longer provided on Linux and Windows. The switch to GLAD allows us to resolve OpenGL symbols when mjr_makeContext is called rather than when the library is loaded. As a result, the MuJoCo library no longer has an explicit dynamic dependency on OpenGL, and can be used on system where OpenGL is not present.

Simulate

4. Fix a bug in simulate where pressing '[' or ']' when a model is not loaded causes a crash.

5. Contact frame visualisation is added to the Simulate GUI.

6. Rename "set key", "reset to key" to "save key" and "load key", respectively.

7. Change bindings of F6 and F7 from the not very useful "vertical sync" and "busy wait" to the more useful cycling of frames and labels.

Bug fixes

8. mj_resetData now zeroes out the solver_nnz field.

9. Remove a special branch in mju_quat2mat for unit quaternions. Previously, mju_quat2mat skipped all computation if the real part of the quaternion equals 1.0. For very small angles (e.g. when finite differencing), the cosine can evaluate to exactly 1.0 at double precision while the sine is still nonzero.

2.1.3

This is a minor release, handling issues discovered after the launch of 2.1.2.

General

1. simulate now support cycling through cameras (with [ and ] keys).
2. mjVIS_STATIC toggles all static bodies, not just direct children of the world.

Python bindings

3. Added a free() method to MjrContext, rather than relying on the object destructor.
4. Enums now support arithmetic and bitwise operations with numbers.

Bug fixes

5. Fixed a rendering bug for planes, introduced in 2.1.2. This broke maze environments in dm_control.