Merge branch 'develop' into antoiner/feat/ovphysx_articulation

Author: kellyguo11

.github/actions/run-tests/action.yml

@@ -214,9 +214,19 @@ runs:
               mkdir -p tests
               rm _isaac_sim || true
               ln -s /isaac-sim _isaac_sim
+              # Allow OmniHub to start in the test container. Some base images
+              # set this detect-only flag, which makes cold asset downloads
+              # fall back to slow repeated retries.
+              unset HUB__ARGS__DETECT_ONLY
               if [ -n \"\${TEST_EXTRA_PIP_PACKAGES:-}\" ]; then
                 echo \"Installing extra pip packages: \${TEST_EXTRA_PIP_PACKAGES}\"
                 ./isaaclab.sh -p -m pip install \${TEST_EXTRA_PIP_PACKAGES}
+                case \" \${TEST_EXTRA_PIP_PACKAGES} \" in
+                  *\" leapp\"*)
+                    echo \"Resolved LEAPP package:\"
+                    ./isaaclab.sh -p -m pip show leapp || true
+                    ;;
+                esac
               fi
               echo 'Starting pytest with path: $test_path'
               ./isaaclab.sh -p -m pytest --ignore=tools/conftest.py --ignore=source/isaaclab/test/install_ci $test_path $pytest_options -v --junitxml=tests/$result_file

CONTRIBUTORS.md

@@ -80,6 +80,7 @@ Guidelines for modifications:
 * Emily Sturman
 * Emmanuel Ferdman
 * Fabian Jenelten
+* Fatima Anes
 * Felipe Mohr
 * Felix Yu
 * Frank Lai
@@ -188,6 +189,7 @@ Guidelines for modifications:
 * Xiaodi Yuan
 * Xinjie Yao
 * Xinpeng Liu
+* Xu Li
 * Yang Jin
 * Yanzi Zhu
 * Yijie Guo

docs/source/experimental-features/bleeding-edge.rst

@@ -11,7 +11,7 @@ To address this, some major features will be released as Experimental Feature Br
 This way, the community can experiment with and contribute to the feature before it's fully integrated, reducing the likelihood of being derailed by unexpected and new errors.
 
 RL Post-Training for VLA Models
----------------------------------
+-------------------------------
 
 `RLinf <https://github.com/RLinf/RLinf.git>`_ is a flexible and scalable open-source RL infrastructure designed for
 Embodied and Agentic AI. This integration enables **reinforcement learning fine-tuning of Vision-Language-Action
@@ -83,6 +83,9 @@ From the Isaac Lab root directory:
    pip install -e .[base] --no-deps
    cd ../
 
+   # Install flash-attn (must be built against the correct PyTorch)
+   pip install --no-build-isolation flash-attn==2.8.3
+
 Quick Start
 ~~~~~~~~~~~
 

docs/source/features/reproducibility.rst

@@ -20,6 +20,29 @@ simulation results are reproducible across different runs. The seed is set into
 parameters :attr:`isaaclab.envs.ManagerBasedEnvCfg.seed` or :attr:`isaaclab.envs.DirectRLEnvCfg.seed`
 depending on the manager-based or direct environment implementation respectively.
 
+App-level deterministic rendering via ``AppLauncher``
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+The ``--deterministic`` flag is provided by :meth:`isaaclab.app.AppLauncher.add_app_launcher_args`.
+After the simulation app starts, :class:`~isaaclab.app.app_launcher.AppLauncher` applies RTX/RTPT carb
+settings via :meth:`~isaaclab.app.app_launcher.AppLauncher.apply_rtx_determinism_settings`.
+
+**Strict PyTorch determinism** (calling :meth:`~isaaclab.utils.seed.configure_seed` with
+``torch_deterministic=True`` when you pass ``--deterministic``) is wired into the RL training scripts
+for **RL-Games**, **skrl**, **RSL-RL**, and **Stable-Baselines3**: each calls
+:meth:`~isaaclab.utils.seed.configure_seed` after constructing its framework runner or agent object
+so library initialization is not disturbed, then training proceeds with the requested global RNG and
+optional PyTorch deterministic algorithms. Whether you need ``--deterministic`` at the app level
+depends on the workload: **physics-only** simulation does not require it; **RTX** rendering
+(non-minimal mode) does require it for reproducible imagery; **Newton** rendering does not require it.
+
+To enable deterministic RTX settings from the app launcher, pass ``--deterministic``.
+
+.. code-block:: bash
+
+  ./isaaclab.sh -p scripts/reinforcement_learning/rl_games/train.py \
+    --task Isaac-Cartpole-RGB-v0 --enable_cameras --headless --deterministic
+
 For results on our determinacy testing for RL training, please check the GitHub Pull Request `#940`_.
 
 .. tip::

docs/source/how-to/import_new_asset.rst

@@ -46,11 +46,12 @@ is then passed to the :class:`~sim.converters.UrdfConverter` class.
 The URDF importer has various configuration parameters that can be set to control the behavior of the importer.
 The default values for the importer's configuration parameters are specified are in the :class:`~sim.converters.UrdfConverterCfg` class, and they are listed below. We made a few commonly modified settings to be available as command-line arguments when calling the ``convert_urdf.py``, and they are marked with ``*`` in the list. For a comprehensive list of the configuration parameters, please check the the documentation at `URDF importer`_.
 
+Articulation and joint structure
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
 * :attr:`~sim.converters.UrdfConverterCfg.fix_base` * - Whether to fix the base of the robot.
   This depends on whether you have a floating-base or fixed-base robot. The command-line flag is
   ``--fix-base`` where when set, the importer will fix the base of the robot, otherwise it will default to floating-base.
-* :attr:`~sim.converters.UrdfConverterCfg.root_link_name` - The link on which the PhysX articulation root is placed.
-  **Deprecated in URDF importer 3.0** — this option is ignored.
 * :attr:`~sim.converters.UrdfConverterCfg.merge_fixed_joints` * - Whether to merge the fixed joints.
   Usually, this should be set to ``True`` to reduce the asset complexity. The command-line flag is
   ``--merge-joints`` where when set, the importer will merge the fixed joints, otherwise it will default to not merging the fixed joints.
@@ -65,10 +66,53 @@ The default values for the importer's configuration parameters are specified are
     We support two ways to set the gains:
 
     * :attr:`~sim.converters.UrdfConverterCfg.JointDriveCfg.PDGainsCfg` - To directly set the stiffness and damping.
+      Both ``stiffness`` and ``damping`` accept a single float (applied uniformly).
     * :attr:`~sim.converters.UrdfConverterCfg.JointDriveCfg.NaturalFrequencyGainsCfg` - To set the gains using the
       desired natural frequency response of the system. **Deprecated in URDF importer 3.0** — use
       ``PDGainsCfg`` instead.
 
+Geometry, collisions, and materials
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+* :attr:`~sim.converters.UrdfConverterCfg.collision_from_visuals` - Whether to create collision geometry
+  from visual geometry when no explicit ``<collision>`` is defined for a link. Defaults to ``False``.
+* :attr:`~sim.converters.UrdfConverterCfg.collision_type` - The collision shape simplification to apply.
+  One of ``"Convex Hull"`` (default), ``"Convex Decomposition"``, ``"Bounding Sphere"``, or ``"Bounding Cube"``.
+* :attr:`~sim.converters.UrdfConverterCfg.self_collision` - Whether to activate self-collisions between
+  links of the articulation. Defaults to ``False``.
+* :attr:`~sim.converters.UrdfConverterCfg.merge_mesh` - Whether to merge meshes where possible to optimize
+  the model. Defaults to ``False``.
+* :attr:`~sim.converters.UrdfConverterCfg.link_density` - Default density in ``kg/m^3`` for links whose
+  ``<inertial>`` properties are missing. ``0.0`` (default) leaves densities unchanged.
+
+Asset resolution and output
+~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+* :attr:`~sim.converters.UrdfConverterCfg.ros_package_paths` - List of ROS package name/path mappings used
+  to resolve ``package://`` URLs in the URDF. Each entry is a dict with keys ``name`` and ``path``.
+* :attr:`~sim.converters.UrdfConverterCfg.robot_type` - Robot type applied by the USD robot schema.
+  Defaults to ``"Default"``. Must be one of: ``"Default"``, ``"End Effector"``, ``"Manipulator"``,
+  ``"Humanoid"``, ``"Wheeled"``, ``"Holonomic"``, ``"Quadruped"``, ``"Mobile Manipulators"``, ``"Aerial"``.
+* :attr:`~sim.converters.UrdfConverterCfg.run_asset_transformer` - Run the asset transformer to convert
+  the flattened USD into a layered USD (interface USD + payloads). Defaults to ``True``.
+* :attr:`~sim.converters.UrdfConverterCfg.run_multi_physics_conversion` - Also emit MuJoCo-compatible joint
+  attributes alongside PhysX. Defaults to ``True``.
+* :attr:`~sim.converters.UrdfConverterCfg.debug_mode` - Write intermediate conversion artifacts next to the
+  output USD for inspection. Defaults to ``False``.
+
+Deprecated (no-op in URDF importer 3.0)
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+The following options are retained for backwards compatibility but are ignored by the URDF importer 3.0.
+A warning is logged when they are set.
+
+* :attr:`~sim.converters.UrdfConverterCfg.root_link_name` - The link on which the PhysX articulation root
+  was previously placed.
+* :attr:`~sim.converters.UrdfConverterCfg.convert_mimic_joints_to_normal_joints` - Convert mimic joints to
+  normal joints during conversion.
+* :attr:`~sim.converters.UrdfConverterCfg.replace_cylinders_with_capsules` - Replace cylinder shapes with
+  capsule shapes during conversion.
+
 For more detailed information on the configuration parameters, please check the documentation for :class:`~sim.converters.UrdfConverterCfg`.
 
 Example Usage
@@ -139,6 +183,14 @@ is derived automatically from the robot name in the URDF):
 
 * ``anymal.usda`` - This is the main asset file.
 
+.. note::
+   The URDF importer auto-deduplicates the per-robot subdirectory when it already exists.
+   If you re-run the converter against the same ``usd_dir`` with a changed configuration
+   (for example, flipping ``fix_base``), the importer writes to a new numbered folder
+   (``anymal_1/``, ``anymal_2/``, …) rather than overwriting the previous output.
+   :attr:`~sim.converters.UrdfConverter.usd_path` reflects whichever folder the importer
+   actually used. Delete stale subdirectories manually (or wipe ``usd_dir``) if you do not
+   want them to accumulate on disk.
 
 To run the script headless, you can add the ``--headless`` flag. This will not open the GUI and
 exit the script after the conversion is complete.
@@ -169,22 +221,64 @@ parameters, please check the the documentation at `MJCF importer`_.
 
 .. note::
    The MJCF importer was rewritten in Isaac Sim 5.0 to use the ``mujoco-usd-converter`` library.
-   Settings such as ``fix_base``, ``import_sites``, ``import_inertia_tensor``, and ``make_instanceable``
-   are no longer needed — the converter now handles these automatically based on the MJCF file content.
+   Settings such as ``import_sites``, ``import_inertia_tensor``, and ``make_instanceable`` are no
+   longer needed — the converter now handles these automatically based on the MJCF file content.
+
+Geometry, collisions, and materials
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
 * :attr:`~sim.converters.MjcfConverterCfg.merge_mesh` * - Whether to merge meshes where possible to
   optimize the model. The command-line flag is ``--merge-mesh``.
 * :attr:`~sim.converters.MjcfConverterCfg.collision_from_visuals` * - Whether to generate collision
   geometry from visual geometries. The command-line flag is ``--collision-from-visuals``.
-* :attr:`~sim.converters.MjcfConverterCfg.collision_type` * - Type of collision geometry to use
-  (e.g. ``"default"``, ``"Convex Hull"``, ``"Convex Decomposition"``). The command-line flag is
-  ``--collision-type``.
+* :attr:`~sim.converters.MjcfConverterCfg.collision_type` * - The collision shape simplification to
+  apply. One of ``"Convex Hull"`` (default), ``"Convex Decomposition"``, ``"Bounding Sphere"``, or
+  ``"Bounding Cube"``. The command-line flag is ``--collision-type``.
 * :attr:`~sim.converters.MjcfConverterCfg.self_collision` * - Whether to activate self-collisions
   between links of the articulation. The command-line flag is ``--self-collision``.
+
+Articulation and physics
+~~~~~~~~~~~~~~~~~~~~~~~~
+
+* :attr:`~sim.converters.MjcfConverterCfg.fix_base` - Whether to add a fixed joint between the world
+  and the root rigid-body link. Defaults to ``False``.
+* :attr:`~sim.converters.MjcfConverterCfg.link_density` - Default density in ``kg/m^3`` for links whose
+  ``<inertial>`` properties are missing in the MJCF. ``0.0`` (default) leaves densities unchanged.
 * :attr:`~sim.converters.MjcfConverterCfg.import_physics_scene` * - Import physics scene properties
   (gravity, time step, etc.) from the MJCF file. Defaults to ``False``. The command-line flag is
   ``--import-physics-scene``.
 
+Actuator overrides
+~~~~~~~~~~~~~~~~~~
+
+MuJoCo models actuators as an affine transformation ``tau = gain @ control + bias``. The following
+options override the values parsed from the MJCF on a per-actuator basis. Each defaults to ``None``,
+which leaves the parsed values unchanged.
+
+* :attr:`~sim.converters.MjcfConverterCfg.override_gain_type` - The actuator gain type override (e.g.
+  ``"fixed"``).
+* :attr:`~sim.converters.MjcfConverterCfg.override_bias_type` - The actuator bias type override (e.g.
+  ``"affine"``).
+* :attr:`~sim.converters.MjcfConverterCfg.override_gain_prm` - The actuator gain parameter array override.
+  Example for position control: ``[kp, 0, 0, 0, 0, 0, 0, 0, 0, 0]``.
+* :attr:`~sim.converters.MjcfConverterCfg.override_bias_prm` - The actuator bias parameter array override.
+  Example for position control: ``[0, -kp, -kd, 0, 0, 0, 0, 0, 0, 0]``.
+
+Asset resolution and output
+~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+* :attr:`~sim.converters.MjcfConverterCfg.robot_type` - Robot type applied by the USD robot schema.
+  Defaults to ``"Default"``. Must be one of: ``"Default"``, ``"End Effector"``, ``"Manipulator"``,
+  ``"Humanoid"``, ``"Wheeled"``, ``"Holonomic"``, ``"Quadruped"``, ``"Mobile Manipulators"``, ``"Aerial"``.
+* :attr:`~sim.converters.MjcfConverterCfg.run_asset_transformer` - Run the asset transformer to convert
+  the flattened USD into a layered USD (interface USD + payloads). Defaults to ``True``.
+* :attr:`~sim.converters.MjcfConverterCfg.run_multi_physics_conversion` - Convert compatible MuJoCo
+  attributes to PhysX attributes (e.g. actuator gains). Defaults to ``True``.
+* :attr:`~sim.converters.MjcfConverterCfg.debug_mode` - Write intermediate conversion artifacts next to
+  the output USD for inspection. Defaults to ``False``.
+
+For more detailed information on the configuration parameters, please check the documentation for :class:`~sim.converters.MjcfConverterCfg`.
+
 
 Example Usage
 ~~~~~~~~~~~~~
@@ -234,6 +328,15 @@ Executing the above script will create the USD file inside the
 
 * ``h1.usd`` - This is the converted USD asset file.
 
+.. note::
+   The MJCF importer auto-deduplicates the per-robot subdirectory when it already exists,
+   matching the URDF importer's behavior. If you re-run the converter against the same
+   ``usd_dir`` with a changed configuration, the importer writes to a new numbered folder
+   (``h1_1/``, ``h1_2/``, …) rather than overwriting the previous output.
+   :attr:`~sim.converters.MjcfConverter.usd_path` reflects whichever folder the importer
+   actually used. Delete stale subdirectories manually (or wipe ``usd_dir``) if you do not
+   want them to accumulate on disk.
+
 .. figure:: ../_static/tutorials/tutorial_convert_mjcf.jpg
     :align: center
     :figwidth: 100%

docs/source/overview/imitation-learning/humanoids_imitation.rst

@@ -636,6 +636,22 @@ Then, from the **Isaac-GR00T** directory, install GR00T N1.5 and its dependencie
    MAX_JOBS=4 uv pip install --no-build-isolation 'git+https://github.com/facebookresearch/pytorch3d.git@v0.7.9'
    uv pip install diffusers decord zmq
 
+.. note::
+
+   **If you cannot install or use flash-attn**, an optional patch is provided that switches the
+   bundled Eagle 2.5 VL model to PyTorch SDPA. Use this if ``flash-attn`` fails to build for your
+   environment, or if it installs but raises a runtime error such as
+   ``RuntimeError: FlashAttention only supports Ampere GPUs or newer`` (for example on Blackwell
+   GPUs, which ``flash-attn==2.7.1.post4`` does not have prebuilt kernels for). After the patch,
+   finetune and rollout run on any CUDA arch supported by your PyTorch build, at the cost of
+   flash-attn's training speedup. Skip the ``flash-attn`` install line above, then apply the
+   patch from the **Isaac-GR00T** directory (the sibling layout above means the IsaacLab
+   checkout is at ``../IsaacLab``):
+
+   .. code:: bash
+
+      git apply ../IsaacLab/scripts/imitation_learning/locomanipulation_sdg/gr00t/no_flash_attn.patch
+
 Convert dataset to LeRobot format
 """""""""""""""""""""""""""""""""
 

docs/source/overview/reinforcement-learning/rl_existing_scripts.rst

@@ -360,15 +360,17 @@ large VLA models that don't fit on a single GPU.
    .. code:: bash
 
       # Step 1: Install RLinf and its dependencies (from isaaclab_contrib)
-      pip install -e "source/isaaclab_contrib[rlinf]"
+      pip install -e "source/isaaclab_contrib[rlinf]" --ignore-requires-python
 
-      # Step 2: Clone and install Isaac-GR00T (for VLA model support)
-      cd scripts/reinforcement_learning/rlinf
+      # Step 2: Clone and install Isaac-GR00T (pinned version, for VLA model support)
       git clone https://github.com/NVIDIA/Isaac-GR00T.git
-      pip install -e Isaac-GR00T/.[base] --no-deps
+      cd Isaac-GR00T
+      git checkout 4af2b622892f7dcb5aae5a3fb70bcb02dc217b96
+      pip install -e .[base] --no-deps
+      cd ../
 
       # Step 3: Install flash-attn (must be built against the correct PyTorch)
-      pip install --no-build-isolation flash-attn==2.7.1.post4
+      pip install --no-build-isolation flash-attn==2.8.3
 
 -  Training a VLA agent with RLinf:
 

docs/source/setup/installation/index.rst

@@ -237,3 +237,4 @@ Please follow the steps :doc:`asset_caching` to enable asset caching and speed u
    source_installation
    isaaclab_pip_installation
    asset_caching
+   uv run (experimental) <uv_run>

docs/source/setup/installation/kitless_installation.rst

@@ -44,7 +44,8 @@ with MJWarp physics and the Newton visualizer:
       .. code-block:: batch
 
          :: Install Isaac Lab (Newton backend, no Isaac Sim required)
-         isaaclab.bat --install   :: or isaaclab.bat -i
+         :: or: isaaclab.bat -i
+         isaaclab.bat --install
 
          :: Kickoff training with MJWarp physics and Newton visualizer
          isaaclab.bat -p scripts\reinforcement_learning\rsl_rl\train.py ^

docs/source/setup/installation/uv_run.rst

@@ -0,0 +1,39 @@
+.. _uv-run-training:
+
+``uv run`` Training and Play (Experimental)
+============================================
+
+.. warning::
+
+   This feature is experimental and subject to change in future releases.
+
+Install ``uv`` if you do not have it already:
+
+.. code-block:: bash
+
+   curl -LsSf https://astral.sh/uv/install.sh | sh
+
+Clone the repo and start training immediately — no virtual environment setup required:
+
+.. code-block:: bash
+
+   git clone https://github.com/isaac-sim/IsaacLab.git
+   cd IsaacLab
+
+   # Newton backend training without Isaac Sim
+   uv run train --rl_library rsl_rl \
+      --task Isaac-Cartpole-Direct-v0 --headless presets=newton_mjwarp
+
+   # Add OVRTX/OVPhysX extras only when the workflow needs them
+   uv run --extra ov --extra rtx train --rl_library rsl_rl \
+      --task Isaac-Cartpole-Direct-v0 --headless presets=newton_mjwarp
+
+``uv`` resolves and manages the environment automatically on each invocation. Supported
+libraries for ``--rl_library`` are: ``rsl_rl``, ``rl_games``, ``skrl``, ``sb3``, and ``rlinf``.
+
+Play / Evaluation
+-----------------
+
+.. code-block:: bash
+
+   uv run play --rl_library rsl_rl --task <any_task>