Version 2.3.3

[saran-t]

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.

---

This discussion was created from the release 2.3.3.

[kevinzakka]

I'm seeing significant speedups in some applications, thank you to the MuJoCo team <3