Add PhysX joint wrench sensor (#5458)

# Description

Adds the PhysX backend for `JointWrenchSensor`.

This PR is built on top of #5412. Until that PR lands,
this comparison also includes the Newton joint-wrench sensor commits
from #5412.

The PhysX sensor creates a PhysX articulation view and reads incoming
joint wrenches from `ArticulationView.get_link_incoming_joint_force()`.
It then converts the raw PhysX body1-frame, body-origin wrench into the
shared `incoming_joint_frame` convention: child-side joint frame with
torque referenced at the child-side joint anchor. PhysX and Newton
joint-wrench sensors also resolve nested USD articulation roots before
creating their articulation views so asset-level paths such as
`{ENV_REGEX_NS}/Robot` work for classic Ant/Humanoid assets.

This also removes `ArticulationData.body_incoming_joint_wrench_b` from
the articulation data APIs, migrates manager/direct environments that
used it to `JointWrenchSensor`, and documents the Isaac Lab 3.0
migration path. Classic Ant/Humanoid Newton presets now use the same
wrench observations as PhysX and pass short RSL-RL training smokes.

@camevor can you review?

Fixes # (issue)

N/A

## Type of change

- New feature (non-breaking change which adds functionality)
- Breaking change (existing functionality will not work without user
modification)
- Documentation update

## Screenshots

N/A

## Tests

- `./isaaclab.sh -f`
- `docker exec isaac-lab-base bash -lc "cd /workspace/isaaclab &&
./isaaclab.sh -p -m pytest
source/isaaclab_physx/test/sensors/test_joint_wrench_sensor.py"` - 11
passed
- `docker exec isaac-lab-base bash -lc "cd /workspace/isaaclab &&
./isaaclab.sh -p -m pytest
source/isaaclab_newton/test/sensors/test_joint_wrench_sensor.py"` - 10
passed
- `docker exec isaac-lab-base bash -lc "cd /workspace/isaaclab &&
./isaaclab.sh -p -m pytest
source/isaaclab_tasks/test/test_environments.py -k 'Isaac-Ant-v0 or
Isaac-Humanoid-v0'"` - 4 passed
- `docker exec isaac-lab-base bash -lc "cd /workspace/isaaclab &&
./isaaclab.sh -p scripts/reinforcement_learning/rsl_rl/train.py --task
Isaac-Ant-v0 --num_envs 16 --max_iterations 1 --headless --device cuda
presets=newton agent.num_steps_per_env=8"` - passed
- `docker exec isaac-lab-base bash -lc "cd /workspace/isaaclab &&
./isaaclab.sh -p scripts/reinforcement_learning/rsl_rl/train.py --task
Isaac-Humanoid-v0 --num_envs 16 --max_iterations 1 --headless --device
cuda presets=newton agent.num_steps_per_env=8"` - passed
- `docker exec isaac-lab-base bash -lc "cd /workspace/isaaclab &&
./isaaclab.sh -p -m pytest
source/isaaclab/test/test_mock_interfaces/test_mock_data_properties.py"`
- 253 passed

## Checklist

- [x] I have read and understood the [contribution
guidelines](https://isaac-sim.github.io/IsaacLab/main/source/refs/contributing.html)
- [x] I have run the [`pre-commit` checks](https://pre-commit.com/) with
`./isaaclab.sh --format`
- [x] I have made corresponding changes to the documentation
- [x] My changes generate no new warnings
- [x] I have added tests that prove my fix is effective or that my
feature works
- [x] I have updated the changelog and the corresponding version in the
extension's `config/extension.toml` file
- [x] I have added my name to the `CONTRIBUTORS.md` or my name already
exists there

---------

Signed-off-by: camevor <camevor@nvidia.com>
Co-authored-by: camevor <camevor@nvidia.com>

Author: AntoineRichard

docs/source/api/lab/isaaclab.sensors.rst

@@ -37,6 +37,7 @@
     Imu
     ImuCfg
     JointWrenchSensor
+    JointWrenchSensorData
     JointWrenchSensorCfg
 
 Sensor Base
@@ -200,6 +201,11 @@ Joint Wrench Sensor
     :inherited-members:
     :show-inheritance:
 
+.. autoclass:: JointWrenchSensorData
+    :members:
+    :inherited-members:
+    :exclude-members: __init__
+
 .. autoclass:: JointWrenchSensorCfg
     :members:
     :inherited-members:

docs/source/migration/migrating_to_isaaclab_3-0.rst

@@ -158,6 +158,8 @@ The following sensor classes also remain in the ``isaaclab`` package with unchan
 - :class:`~isaaclab.sensors.Imu`, :class:`~isaaclab.sensors.ImuCfg`, :class:`~isaaclab.sensors.ImuData`
 - :class:`~isaaclab.sensors.Pva`, :class:`~isaaclab.sensors.PvaCfg`, :class:`~isaaclab.sensors.PvaData`
 - :class:`~isaaclab.sensors.FrameTransformer`, :class:`~isaaclab.sensors.FrameTransformerCfg`, :class:`~isaaclab.sensors.FrameTransformerData`
+- :class:`~isaaclab.sensors.JointWrenchSensor`, :class:`~isaaclab.sensors.JointWrenchSensorCfg`,
+  :class:`~isaaclab.sensors.JointWrenchSensorData`
 
 These sensor classes now use factory patterns that automatically instantiate the appropriate backend
 implementation (PhysX by default), maintaining full backward compatibility.
@@ -179,6 +181,7 @@ you can import from ``isaaclab_physx.sensors``:
    from isaaclab_physx.sensors import Imu, ImuData
    from isaaclab_physx.sensors import Pva, PvaData
    from isaaclab_physx.sensors import FrameTransformer, FrameTransformerData
+   from isaaclab_physx.sensors import JointWrenchSensor, JointWrenchSensorData
 
 
 New ``isaaclab_newton`` Extension
@@ -188,6 +191,8 @@ A new extension ``isaaclab_newton`` provides Newton physics backend implementati
 
 - :class:`~isaaclab_newton.assets.Articulation` and :class:`~isaaclab_newton.assets.ArticulationData`
 - :class:`~isaaclab_newton.assets.RigidObject` and :class:`~isaaclab_newton.assets.RigidObjectData`
+- :class:`~isaaclab_newton.sensors.JointWrenchSensor` and
+  :class:`~isaaclab_newton.sensors.JointWrenchSensorData`
 
 These classes implement the same base interfaces as their PhysX counterparts
 (:class:`~isaaclab.assets.BaseArticulation`, :class:`~isaaclab.assets.BaseRigidObject`),
@@ -331,6 +336,75 @@ If you need to track sensor poses in world frame, please use a dedicated sensor
    sensor_quat = frame_transformer.data.target_quat_w
 
 
+Articulation Joint Wrench Data Moved to ``JointWrenchSensor``
+-------------------------------------------------------------
+
+The ``ArticulationData.body_incoming_joint_wrench_b`` property has been removed. In
+Isaac Lab 3.0, incoming joint reaction wrenches are exposed through
+:class:`~isaaclab.sensors.JointWrenchSensor`, which has PhysX and Newton backend
+implementations and returns separate force [N] and torque [N·m] buffers.
+The sensor reports wrenches in the child-side incoming joint frame, with torque
+referenced at the child-side joint anchor.
+
+**Before (Isaac Lab 2.x):**
+
+.. code-block:: python
+
+   wrench_b = robot.data.body_incoming_joint_wrench_b.torch[:, body_ids]
+
+**After (Isaac Lab 3.x):**
+
+.. code-block:: python
+
+   import torch
+   from isaaclab.scene import InteractiveSceneCfg
+   from isaaclab.sensors import JointWrenchSensorCfg
+
+   class MySceneCfg(InteractiveSceneCfg):
+       robot = ROBOT_CFG.replace(prim_path="{ENV_REGEX_NS}/Robot")
+       joint_wrench = JointWrenchSensorCfg(prim_path="{ENV_REGEX_NS}/Robot")
+
+   sensor = env.scene.sensors["joint_wrench"]
+   data = sensor.data
+   wrench_j = torch.cat(
+       (
+           data.force.torch[:, body_ids],
+           data.torque.torch[:, body_ids],
+       ),
+       dim=-1,
+   )
+
+Use :attr:`~isaaclab.sensors.BaseJointWrenchSensor.body_names` or
+:meth:`~isaaclab.sensors.BaseJointWrenchSensor.find_bodies` to map sensor entries to
+articulation body names. PhysX reports one entry for every link, including the articulation
+root link. Newton reports the child bodies of reportable incoming joints.
+
+For manager-based environments, update observations that used the articulation data property to
+depend on the joint-wrench sensor instead:
+
+.. code-block:: python
+
+   import isaaclab.envs.mdp as mdp
+   from isaaclab.managers import SceneEntityCfg
+   from isaaclab.managers import ObservationTermCfg as ObsTerm
+   from isaaclab.scene import InteractiveSceneCfg
+   from isaaclab.sensors import JointWrenchSensorCfg
+
+   class MySceneCfg(InteractiveSceneCfg):
+       robot = ROBOT_CFG.replace(prim_path="{ENV_REGEX_NS}/Robot")
+       joint_wrench = JointWrenchSensorCfg(prim_path="{ENV_REGEX_NS}/Robot")
+
+   feet_body_forces = ObsTerm(
+       func=mdp.body_incoming_wrench,
+       params={
+           "sensor_cfg": SceneEntityCfg(
+               "joint_wrench",
+               body_names=["left_foot", "right_foot"],
+           )
+       },
+   )
+
+
 Multi-Backend Support: PresetCfg Pattern
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 

source/isaaclab/changelog.d/pr-5458-merge-develop.rst

@@ -0,0 +1,15 @@
+Changed
+^^^^^^^
+
+* Changed :func:`~isaaclab.envs.mdp.body_incoming_wrench` to read from
+  :class:`~isaaclab.sensors.JointWrenchSensor`. Pass
+  ``sensor_cfg=SceneEntityCfg("joint_wrench", body_names=...)`` instead of an
+  articulation asset config.
+
+Removed
+^^^^^^^
+
+* Removed ``BaseArticulationData.body_incoming_joint_wrench_b``. Add
+  :class:`~isaaclab.sensors.JointWrenchSensorCfg` to the scene and read
+  :attr:`~isaaclab.sensors.JointWrenchSensorData.force` and
+  :attr:`~isaaclab.sensors.JointWrenchSensorData.torque` instead.

source/isaaclab/docs/CHANGELOG.rst

@@ -1,7 +1,6 @@
 Changelog
 ---------
 
-
 4.6.27 (2026-05-01)
 ~~~~~~~~~~~~~~~~~~~
 

source/isaaclab/isaaclab/assets/articulation/base_articulation_data.py

@@ -664,24 +664,6 @@ def body_com_pose_b(self) -> ProxyArray:
         """
         raise NotImplementedError
 
-    @property
-    @abstractmethod
-    @leapp_tensor_semantics(kind=InputKindEnum.WRENCH)
-    def body_incoming_joint_wrench_b(self) -> ProxyArray:
-        """Joint reaction wrench applied from body parent to child body in parent body frame.
-
-        Shape is (num_instances, num_bodies), dtype = wp.spatial_vectorf. In torch this resolves to
-        (num_instances, num_bodies, 6). All body reaction wrenches are provided including the root body to the
-        world of an articulation.
-
-        For more information on joint wrenches, please check the `PhysX documentation`_ and the
-        underlying `PhysX Tensor API`_.
-
-        .. _PhysX documentation: https://nvidia-omniverse.github.io/PhysX/physx/5.5.1/docs/Articulations.html#link-incoming-joint-force
-        .. _PhysX Tensor API: https://docs.omniverse.nvidia.com/kit/docs/omni_physics/latest/extensions/runtime/source/omni.physics.tensors/docs/api/python.html#omni.physics.tensors.api.ArticulationView.get_link_incoming_joint_force
-        """
-        raise NotImplementedError
-
     ##
     # Joint state properties.
     ##

source/isaaclab/isaaclab/envs/mdp/observations.py

@@ -23,7 +23,7 @@
 if TYPE_CHECKING:
     from isaaclab.assets import Articulation, RigidObject
     from isaaclab.envs import ManagerBasedEnv, ManagerBasedRLEnv
-    from isaaclab.sensors import Camera, Imu, Pva, RayCaster, RayCasterCamera
+    from isaaclab.sensors import Camera, Imu, JointWrenchSensor, Pva, RayCaster, RayCasterCamera
 
 from isaaclab.envs.utils.io_descriptors import (
     generic_io_descriptor,
@@ -304,16 +304,21 @@ def height_scan(env: ManagerBasedEnv, sensor_cfg: SceneEntityCfg, offset: float
     return sensor.data.pos_w.torch[:, 2].unsqueeze(1) - sensor.data.ray_hits_w.torch[..., 2] - offset
 
 
-def body_incoming_wrench(env: ManagerBasedEnv, asset_cfg: SceneEntityCfg) -> torch.Tensor:
-    """Incoming spatial wrench on bodies of an articulation in the simulation world frame.
+def body_incoming_wrench(env: ManagerBasedEnv, sensor_cfg: SceneEntityCfg) -> torch.Tensor:
+    """Incoming spatial wrench [N, N·m] on bodies of an articulation in the sensor convention.
 
-    This is the 6-D wrench (force and torque) applied to the body link by the incoming joint force.
+    This is the 6-D wrench (force followed by torque) applied to the body link by the incoming joint force.
     """
     # extract the used quantities (to enable type-hinting)
-    asset: Articulation = env.scene[asset_cfg.name]
-    # obtain the link incoming forces in world frame
-    body_incoming_joint_wrench_b = asset.data.body_incoming_joint_wrench_b.torch[:, asset_cfg.body_ids]
-    return body_incoming_joint_wrench_b.view(env.num_envs, -1)
+    sensor: JointWrenchSensor = env.scene.sensors[sensor_cfg.name]
+    sensor_data = sensor.data
+    force_data = sensor_data.force
+    torque_data = sensor_data.torque
+    if force_data is None or torque_data is None:
+        raise RuntimeError("Joint wrench sensor data is not initialized. Call sim.reset() before reading observations.")
+    force = force_data.torch[:, sensor_cfg.body_ids]
+    torque = torque_data.torch[:, sensor_cfg.body_ids]
+    return torch.cat((force, torque), dim=-1).view(env.num_envs, -1)
 
 
 def pva_orientation(env: ManagerBasedEnv, asset_cfg: SceneEntityCfg = SceneEntityCfg("pva")) -> torch.Tensor:

source/isaaclab/isaaclab/sensors/joint_wrench/base_joint_wrench_sensor.py

@@ -6,10 +6,13 @@
 from __future__ import annotations
 
 from abc import abstractmethod
+from collections.abc import Sequence
 from typing import TYPE_CHECKING
 
 import warp as wp
 
+import isaaclab.utils.string as string_utils
+
 from ..sensor_base import SensorBase
 from .base_joint_wrench_sensor_data import BaseJointWrenchSensorData
 
@@ -20,11 +23,12 @@
 class BaseJointWrenchSensor(SensorBase):
     """The joint reaction wrench sensor.
 
-    Reports the incoming joint wrench on each joint of an articulation as a
-    split force [N] / torque [N·m] pair expressed in the
+    Reports incoming joint wrenches for the bodies selected by the backend as
+    split force [N] / torque [N·m] pairs expressed in the
     ``INCOMING_JOINT_FRAME`` convention (child-side joint frame, child-side
     joint anchor reference point). Backends convert from their native
-    representation to this convention internally.
+    representation to this convention internally. Use :attr:`body_names` or
+    :meth:`find_bodies` to map entries to articulation bodies.
     """
 
     cfg: JointWrenchSensorCfg
@@ -57,6 +61,27 @@ def body_names(self) -> list[str]:
         """Ordered names of the bodies whose incoming joint wrench is reported."""
         raise NotImplementedError
 
+    @property
+    def num_bodies(self) -> int:
+        """Number of bodies whose incoming joint wrench is reported."""
+        return len(self.body_names)
+
+    """
+    Operations
+    """
+
+    def find_bodies(self, name_keys: str | Sequence[str], preserve_order: bool = False) -> tuple[list[int], list[str]]:
+        """Find reported bodies based on name keys.
+
+        Args:
+            name_keys: A regular expression or list of regular expressions to match the body names.
+            preserve_order: Whether to preserve the order of the name keys in the output. Defaults to False.
+
+        Returns:
+            The matching body indices and names.
+        """
+        return string_utils.resolve_matching_names(name_keys, self.body_names, preserve_order)
+
     """
     Implementation - Abstract methods to be implemented by backend-specific subclasses.
     """

source/isaaclab/isaaclab/sensors/joint_wrench/base_joint_wrench_sensor_data.py

@@ -22,8 +22,8 @@ def force(self) -> ProxyArray | None:
 
         Expressed in the frame selected by
         :attr:`~isaaclab.sensors.JointWrenchSensorCfg.convention`. Shape is
-        ``(num_envs, num_joints)``, dtype ``wp.vec3f``. In torch this resolves
-        to ``(num_envs, num_joints, 3)``. ``None`` before the simulation is
+        ``(num_envs, num_bodies)``, dtype ``wp.vec3f``. In torch this resolves
+        to ``(num_envs, num_bodies, 3)``. ``None`` before the simulation is
         initialized.
         """
         raise NotImplementedError
@@ -35,8 +35,8 @@ def torque(self) -> ProxyArray | None:
 
         Expressed in the frame selected by
         :attr:`~isaaclab.sensors.JointWrenchSensorCfg.convention`. Shape is
-        ``(num_envs, num_joints)``, dtype ``wp.vec3f``. In torch this resolves
-        to ``(num_envs, num_joints, 3)``. ``None`` before the simulation is
+        ``(num_envs, num_bodies)``, dtype ``wp.vec3f``. In torch this resolves
+        to ``(num_envs, num_bodies, 3)``. ``None`` before the simulation is
         initialized.
         """
         raise NotImplementedError

source/isaaclab/isaaclab/sensors/joint_wrench/joint_wrench_sensor.py

@@ -15,13 +15,15 @@
 if TYPE_CHECKING:
     from isaaclab_newton.sensors.joint_wrench import JointWrenchSensor as NewtonJointWrenchSensor
     from isaaclab_newton.sensors.joint_wrench import JointWrenchSensorData as NewtonJointWrenchSensorData
+    from isaaclab_physx.sensors.joint_wrench import JointWrenchSensor as PhysXJointWrenchSensor
+    from isaaclab_physx.sensors.joint_wrench import JointWrenchSensorData as PhysXJointWrenchSensorData
 
 
 class JointWrenchSensor(FactoryBase, BaseJointWrenchSensor):
     """Factory for creating joint-wrench sensor instances."""
 
-    data: BaseJointWrenchSensorData | NewtonJointWrenchSensorData
+    data: BaseJointWrenchSensorData | PhysXJointWrenchSensorData | NewtonJointWrenchSensorData
 
-    def __new__(cls, *args, **kwargs) -> BaseJointWrenchSensor | NewtonJointWrenchSensor:
+    def __new__(cls, *args, **kwargs) -> BaseJointWrenchSensor | PhysXJointWrenchSensor | NewtonJointWrenchSensor:
         """Create a new instance of a joint-wrench sensor based on the backend."""
         return super().__new__(cls, *args, **kwargs)

source/isaaclab/isaaclab/sensors/joint_wrench/joint_wrench_sensor_cfg.py

@@ -25,6 +25,6 @@ class JointWrenchSensorCfg(SensorBaseCfg):
     """Coordinate convention for the reported wrench. Defaults to ``"incoming_joint_frame"``.
 
     - ``"incoming_joint_frame"`` — child-side joint frame, child-side joint anchor as reference point.
-      Matches what a real 6-axis F/T sensor mounted at the joint would measure. This is the same
-      as PhysX convention in IsaacLab2.3
+      Matches what a real 6-axis F/T sensor mounted at the joint would measure. Backends convert
+      their native solver outputs to this convention.
     """