EnvPool MJLab implementation
===========================

The native task managers and serialized model assets are derived from MJLab
1.6.0, commit b517e0c489139e7fcee95702cfb2b01931264985, distributed under the
Apache License, Version 2.0. The pinned source includes the Unitree G1, Unitree
Go1 and I2RT Yam models used by these presets. Original upstream license files
are included with the runtime assets.

The quaternion expressions in math.h follow MJLab's Isaac Lab math helpers.
Those helpers retain the Isaac Lab Developers' BSD-3-Clause attribution; see
LICENSE.isaaclab. Random-number conventions follow PyTorch 2.9.0, commit
0fabc3ba44823f257e70ce397d989c8de5e362c1; see LICENSE.pytorch. The native
implementation uses standard C++ math and MuJoCo's small-matrix routines for
reward calculations, without a dependency on Torch at runtime.

Physics and observation-camera kernels are generated from MuJoCo-Warp 3.11.0;
see LICENSE.mujoco-warp and AUTHORS.mujoco-warp for its Apache-2.0 license and
attribution.
The CPU runtime and generated kernel support use NVIDIA Warp 1.14.0, commit
b943176fbe3ab70e90708e74d2c48e6f50557145. Warp is distributed under the Apache
License, Version 2.0 and incorporates third-party components with their own
notices. The Warp license and its supplied third-party notices are preserved
in the vendor license directory. EnvPool's changes to Warp are recorded in
warp_native.patch; they add static CPU replay and thread-local scratch storage.

The locally generated motion fixture is for testing only. It is not included
in either the runtime asset package or the EnvPool wheel.
