MuJoCo 3.5

[yuvaltassa]

3.5.0 is a significant release. Please see the changelog for a full list of changes. Below are some of the highlights:

1. MuJoCo Warp (MJWarp) is now officially released.

   MJWarp is a reimplementation of MuJoCo written in Warp, optimized for NVIDIA GPUs and massively parallel simulation. It is developed and maintained as a joint effort by NVIDIA and Google DeepMind.

2. Added a new System Identification toolbox (Python), see the README for details.
   A Colab notebook demonstrating the toolbox is included.
   Contribution by @kevinzakka, @aftersomemath, @jonathanembleyriches, @qiayuanl, @spjardim and @gizemozd.

3. Actuators and sensors now support arbitrary delays via history buffers, and sensor values can be computed at intervals larger than the simulation timestep. Using a delay or interval introduces a new mjData.history variable to the Physics state. See Delays for details.

4. Added new flexvert equality constraints that enable cloth simulations with coarser meshes, using the method described in Chen, Kry and Vouga, 2019.

5. Added implicit integration support for deformable objects (flex) in implicit and implicitfast integrators. This method extracts the flex degrees of freedom and solves them as a dense block, enabling increased stability for stiff flex objects without reducing the timestep. Compatible with the trilinear and quadratic dof types.

6. Rangefinder sensors can now be attached to a camera, in which case the sensor respects the camera resolution and casts multiple rays, one for each pixel.

7. Rangefinders can now report various kinds of information beyond ray distances, including surface normals and intersection points.

[HoangGiang93]

Amazing sys-id toolbox!