[Task Clean-up][Manager] Dexterous Part 5/9: Add the reorientation manager counterparts

docs/source/setup/quickstart_details.rst

@@ -57,7 +57,7 @@ options (observation modes, camera configs, etc.). They fold into Hydra override
 
          # OVRTX rendering (kit-less, no Kit visualizer)
          uv run isaaclab train --rl_library rsl_rl \
-           --task=Isaac-Reorient-Cube-Shadow-Camera-Benchmark-Direct \
+           --task=IsaacContrib-Reorient-Cube-Shadow-Camera-Benchmark-Direct \
            --num_envs=16 --max_iterations=10 \
            physics=newton_mjwarp renderer=ovrtx presets=simple_shading_diffuse_mdl
 

source/isaaclab/changelog.d/task-cleanup-dex-part08.rst

@@ -0,0 +1,6 @@
+Fixed
+^^^^^
+
+* Fixed :attr:`~isaaclab.envs.ManagerBasedRLEnv.reset_buf` not existing until the first
+  call to :meth:`~isaaclab.envs.ManagerBasedRLEnv.step`, so manager terms that run during
+  the initial reset could not read it.

source/isaaclab/isaaclab/envs/manager_based_rl_env.py

@@ -73,8 +73,11 @@ def __init__(self, cfg: ManagerBasedRLEnvCfg, render_mode: str | None = None, **
         # -- counter for curriculum
         self.common_step_counter = 0
 
-        # initialize the episode length buffer BEFORE loading the managers to use it in mdp functions.
+        # initialize the episode length and reset buffers BEFORE loading the managers to use them in
+        # mdp functions. The reset buffer is only assigned its computed value in :meth:`step`, so terms
+        # that run during the initial reset would otherwise not find it.
         self.episode_length_buf = torch.zeros(cfg.scene.num_envs, device=cfg.sim.device, dtype=torch.long)
+        self.reset_buf = torch.zeros(cfg.scene.num_envs, device=cfg.sim.device, dtype=torch.bool)
 
         # Forward render_mode and viewer camera to VideoRecorderCfg before super().__init__()
         # creates the VideoRecorder, so fallback cameras are only spawned when --video is active

source/isaaclab_tasks/changelog.d/task-cleanup-dex-part08.major.rst

@@ -0,0 +1,74 @@
+Added
+^^^^^
+
+* Added manager-based counterparts for the Shadow cube reorientation task and
+  its OpenAI FF/LSTM observation variants, alongside the existing Allegro
+  manager task.
+* Added :class:`~isaaclab_tasks.core.reorient.mdp.reorient_timeout`, which
+  restarts the episode timer on every goal reach so OpenAI-variant episodes
+  extend across success streaks.
+* Added ``enable_domain_randomization`` to the manager-based Allegro and Shadow
+  environment for turning off its startup randomization terms.
+* Added Newton and OvPhysx physics presets to the manager-based reorientation
+  environments, selectable with ``physics=``.
+* Added a Direct-versus-manager value-parity check covering timing, success
+  tolerance, fall distance, and the consecutive-success cap.
+
+Changed
+^^^^^^^
+
+* **Breaking:** Changed the manager-based reorientation tasks to report
+  ``Metrics/success_rate`` as a per-episode success bit, drawn at
+  ``ReorientCommandCfg.success_count_threshold`` like the Direct environments,
+  instead of a per-attempt ratio. Curves from earlier runs are not comparable.
+* **Breaking:** Changed domain randomization to default off on the Allegro and
+  Shadow manager tasks, so they match their Direct counterparts, which randomize
+  nothing beyond the reset distributions. Set ``enable_domain_randomization`` on
+  the configuration to restore it. The OpenAI variants keep it enabled, matching
+  their Direct counterparts.
+* **Breaking:** Changed the manager-based Allegro task to use the same RL agent
+  configurations as its Direct counterpart, so the two are comparable. The
+  ``rl_games_manager_ppo_cfg.yaml``, ``skrl_manager_ppo_cfg.yaml`` and
+  ``AllegroCubePPORunnerCfg`` entries are removed; use ``rl_games_ppo_cfg.yaml``,
+  ``skrl_ppo_cfg.yaml`` and ``AllegroHandPPORunnerCfg``.
+
+* **Breaking:** Changed the manager-based Allegro reorientation environment to
+  match the Direct observation, action, reset, and termination contracts. The
+  observation space changes size, so existing manager checkpoints cannot be
+  loaded and must be retrained.
+* **Breaking:** Changed the Shadow Hand reorientation tasks to apply the same
+  randomization on every physics backend. PhysX now also randomizes joint gains,
+  object mass, and gravity, and Newton now also randomizes contact materials.
+  Policies trained before this change must be retrained.
+* **Breaking:** Moved the Shadow Hand camera benchmark task to the contributed
+  tasks as ``IsaacContrib-Reorient-Cube-Shadow-Camera-Benchmark-Direct``. The
+  released ``Isaac-Reorient-Cube-Shadow-Camera-Benchmark-Direct`` identifier no
+  longer resolves; use the contributed identifier instead.
+* **Breaking:** Changed the in-hand cube to a rigid body on the Newton backend,
+  where it was previously declared as an articulation with no joints or
+  actuators. Code that resolved the object through
+  :class:`~isaaclab.assets.Articulation` must use
+  :class:`~isaaclab.assets.RigidObject`.
+* Removed the ``clone_in_fabric`` settings from the reorientation scenes. The
+  flag no longer reaches the replicator, so the value had no effect.
+* Renamed the per-robot scene constants to name what they hold: ``ROBOT_CFG``
+  becomes ``SHADOW_HAND_ROBOT_CFG`` or ``ALLEGRO_HAND_ROBOT_CFG``,
+  ``OBJECT_CFG`` becomes ``CUBE_CFG``, and ``ObjectCfg`` becomes ``CubeCfg``.
+
+Removed
+^^^^^^^
+
+* Removed ``ReorientObjectEnvCfg`` and the shared reorientation observation,
+  action, and command configurations. Each manager task now declares its own;
+  derive from :class:`~isaaclab.envs.ManagerBasedRLEnvCfg` directly.
+* Removed ``reorient_common``. Its constants are declared by the tasks that use
+  them, and the in-hand offset and goal-marker position are now per-robot fields
+  on the Direct configurations.
+
+Fixed
+^^^^^
+
+* Fixed the manager-based reorientation tasks not reporting
+  ``Metrics/success_rate``.
+* Fixed manager ``Metrics/success_rate`` counting goal attempts rather than the
+  per-episode success bit the Direct tasks report.

source/isaaclab_tasks/isaaclab_tasks/contrib/reorient/__init__.py

@@ -0,0 +1,6 @@
+# Copyright (c) 2022-2026, The Isaac Lab Project Developers (https://github.com/isaac-sim/IsaacLab/blob/main/CONTRIBUTORS.md).
+# All rights reserved.
+#
+# SPDX-License-Identifier: BSD-3-Clause
+
+"""Contributed variants of the in-hand reorientation tasks."""

source/isaaclab_tasks/isaaclab_tasks/contrib/reorient/config/__init__.py

@@ -0,0 +1,6 @@
+# Copyright (c) 2022-2026, The Isaac Lab Project Developers (https://github.com/isaac-sim/IsaacLab/blob/main/CONTRIBUTORS.md).
+# All rights reserved.
+#
+# SPDX-License-Identifier: BSD-3-Clause
+
+"""Configurations for the contributed reorientation environments."""

source/isaaclab_tasks/isaaclab_tasks/contrib/reorient/config/shadow_hand/__init__.py

@@ -0,0 +1,21 @@
+# Copyright (c) 2022-2026, The Isaac Lab Project Developers (https://github.com/isaac-sim/IsaacLab/blob/main/CONTRIBUTORS.md).
+# All rights reserved.
+#
+# SPDX-License-Identifier: BSD-3-Clause
+
+"""Shadow Hand rendering-throughput benchmark task."""
+
+import gymnasium as gym
+
+from isaaclab_tasks.core.reorient.config.shadow_hand import agents
+
+gym.register(
+    id="IsaacContrib-Reorient-Cube-Shadow-Camera-Benchmark-Direct",
+    entry_point="isaaclab_tasks.core.reorient.config.shadow_hand.shadow_hand_direct_camera_env:ShadowHandCameraEnv",
+    disable_env_checker=True,
+    kwargs={
+        "env_cfg_entry_point": f"{__name__}.shadow_hand_camera_benchmark_env_cfg:ShadowHandCameraBenchmarkEnvCfg",
+        "rsl_rl_cfg_entry_point": f"{agents.__name__}.rsl_rl_ppo_cfg:ShadowHandCameraFFPPORunnerCfg",
+        "rl_games_cfg_entry_point": f"{agents.__name__}:rl_games_ppo_camera_cfg.yaml",
+    },
+)

source/isaaclab_tasks/isaaclab_tasks/contrib/reorient/config/shadow_hand/shadow_hand_camera_benchmark_env_cfg.py

@@ -0,0 +1,30 @@
+# Copyright (c) 2022-2026, The Isaac Lab Project Developers (https://github.com/isaac-sim/IsaacLab/blob/main/CONTRIBUTORS.md).
+# All rights reserved.
+#
+# SPDX-License-Identifier: BSD-3-Clause
+
+"""Rendering-throughput benchmark variant of the Shadow Hand camera task."""
+
+from isaaclab.utils.configclass import configclass
+
+from isaaclab_tasks.core.reorient.config.shadow_hand.feature_extractor import FeatureExtractorCfg
+from isaaclab_tasks.core.reorient.config.shadow_hand.shadow_hand_direct_camera_env_cfg import ShadowHandCameraEnvCfg
+
+
+@configclass
+class ShadowHandCameraBenchmarkEnvCfg(ShadowHandCameraEnvCfg):
+    """Benchmark configuration with the feature extractor CNN disabled.
+
+    The tiled camera renders frames each step as normal, but the CNN forward pass is
+    bypassed — zero embeddings are returned instead. This isolates rendering throughput
+    from CNN inference overhead when profiling.
+
+    The renderer backend and camera data types can still be selected via ``presets``::
+
+        presets = newton_renderer  # benchmark with Newton renderer
+        presets = ovrtx  # benchmark with OVRTX renderer
+        presets = rgb  # benchmark RGB rendering only
+        presets = depth, newton_renderer  # benchmark depth rendering with Newton
+    """
+
+    feature_extractor: FeatureExtractorCfg = FeatureExtractorCfg(enabled=False)

source/isaaclab_tasks/isaaclab_tasks/core/handover/handover_env_cfg.py

@@ -8,13 +8,10 @@
 from isaaclab_ovphysx.physics import OvPhysxCfg
 from isaaclab_physx.physics import PhysxCfg
 
-import isaaclab.envs.mdp as mdp
 import isaaclab.sim as sim_utils
 import isaaclab.utils.math as math_utils
 from isaaclab.assets import ArticulationCfg, RigidObjectCfg
 from isaaclab.envs import DirectMARLEnvCfg
-from isaaclab.managers import EventTermCfg as EventTerm
-from isaaclab.managers import SceneEntityCfg
 from isaaclab.markers import VisualizationMarkersCfg
 from isaaclab.physics import PhysxAutoCfg
 from isaaclab.scene import InteractiveSceneCfg
@@ -33,105 +30,6 @@
 from isaaclab_assets.robots.shadow_hand import SHADOW_HAND_CFG, SHADOW_HAND_NEWTON_CFG
 
 
-@configclass
-class EventCfg:
-    """Configuration for randomization (PhysX path).
-
-    Note: this config is currently not wired into ``HandoverEnvCfg.events`` -
-    it is kept as a reference for future event-randomization work. The event
-    terms here use PhysX-only APIs (rigid-body materials, fixed tendons), so
-    they would need a Newton variant before being enabled in the env.
-    """
-
-    # -- robot
-    robot_physics_material = EventTerm(
-        func=mdp.randomize_rigid_body_material,
-        mode="reset",
-        min_step_count_between_reset=720,
-        params={
-            "asset_cfg": SceneEntityCfg("right_hand"),
-            "static_friction_range": (0.7, 1.3),
-            "dynamic_friction_range": (1.0, 1.0),
-            "restitution_range": (1.0, 1.0),
-            "num_buckets": 250,
-        },
-    )
-    robot_joint_stiffness_and_damping = EventTerm(
-        func=mdp.randomize_actuator_gains,
-        min_step_count_between_reset=720,
-        mode="reset",
-        params={
-            "asset_cfg": SceneEntityCfg("right_hand", joint_names=".*"),
-            "stiffness_distribution_params": (0.75, 1.5),
-            "damping_distribution_params": (0.3, 3.0),
-            "operation": "scale",
-            "distribution": "log_uniform",
-        },
-    )
-    robot_joint_pos_limits = EventTerm(
-        func=mdp.randomize_joint_parameters,
-        min_step_count_between_reset=720,
-        mode="reset",
-        params={
-            "asset_cfg": SceneEntityCfg("right_hand", joint_names=".*"),
-            "lower_limit_distribution_params": (0.00, 0.01),
-            "upper_limit_distribution_params": (0.00, 0.01),
-            "operation": "add",
-            "distribution": "gaussian",
-        },
-    )
-    robot_tendon_properties = EventTerm(
-        func=mdp.randomize_fixed_tendon_parameters,
-        min_step_count_between_reset=720,
-        mode="reset",
-        params={
-            "asset_cfg": SceneEntityCfg("right_hand", fixed_tendon_names=".*"),
-            "stiffness_distribution_params": (0.75, 1.5),
-            "damping_distribution_params": (0.3, 3.0),
-            "operation": "scale",
-            "distribution": "log_uniform",
-        },
-    )
-
-    # -- object
-    object_physics_material = EventTerm(
-        func=mdp.randomize_rigid_body_material,
-        min_step_count_between_reset=720,
-        mode="reset",
-        params={
-            "asset_cfg": SceneEntityCfg("object"),
-            "static_friction_range": (0.7, 1.3),
-            "dynamic_friction_range": (1.0, 1.0),
-            "restitution_range": (1.0, 1.0),
-            "num_buckets": 250,
-        },
-    )
-    object_scale_mass = EventTerm(
-        func=mdp.randomize_rigid_body_mass,
-        min_step_count_between_reset=720,
-        mode="reset",
-        params={
-            "asset_cfg": SceneEntityCfg("object"),
-            "mass_distribution_params": (0.5, 1.5),
-            "operation": "scale",
-            "distribution": "uniform",
-        },
-    )
-
-    # -- scene
-    reset_gravity = EventTerm(
-        func=mdp.randomize_physics_scene_gravity,
-        mode="interval",
-        is_global_time=True,
-        interval_range_s=(36.0, 36.0),  # time_s = num_steps * (decimation * dt)
-        params={
-            "gravity_distribution_params": ([0.0, 0.0, 0.0], [0.0, 0.0, 0.4]),
-            "operation": "add",
-            "distribution": "gaussian",
-        },
-    )
-
-
 def _shadow_hand_cfg(
     prim_path: str,
     init_pos: tuple[float, float, float],

source/isaaclab_tasks/isaaclab_tasks/core/reorient/__init__.py

@@ -8,10 +8,9 @@
 This package consolidates the direct-workflow and manager-based-workflow in-hand
 manipulation tasks, where a dexterous hand reorients an object to match a goal
 orientation. The shared direct base environment lives in
-:mod:`~isaaclab_tasks.core.reorient.reorient_direct_env` and the shared manager-based
-base configuration in :mod:`~isaaclab_tasks.core.reorient.reorient_manager_env_cfg`.
-Robot-specific tasks are organized under the ``config`` subpackage
-(``config/allegro_hand`` and ``config/shadow_hand``).
+:mod:`~isaaclab_tasks.core.reorient.reorient_direct_env`; the manager-based
+configurations live with their robot-specific tasks under the ``config``
+subpackage (``config/allegro_hand`` and ``config/shadow_hand``).
 
 These environments are based on the `dexterous cube manipulation`_ environments
 provided in IsaacGymEnvs repository from NVIDIA. However, they contain certain

source/isaaclab_tasks/isaaclab_tasks/core/reorient/config/allegro_hand/__init__.py

@@ -19,9 +19,9 @@
     disable_env_checker=True,
     kwargs={
         "env_cfg_entry_point": f"{__name__}.allegro_hand_manager_env_cfg:AllegroCubeEnvCfg",
-        "rsl_rl_cfg_entry_point": f"{agents.__name__}.rsl_rl_ppo_cfg:AllegroCubePPORunnerCfg",
-        "rl_games_cfg_entry_point": f"{agents.__name__}:rl_games_manager_ppo_cfg.yaml",
-        "skrl_cfg_entry_point": f"{agents.__name__}:skrl_manager_ppo_cfg.yaml",
+        "rsl_rl_cfg_entry_point": f"{agents.__name__}.rsl_rl_ppo_cfg:AllegroHandPPORunnerCfg",
+        "rl_games_cfg_entry_point": f"{agents.__name__}:rl_games_ppo_cfg.yaml",
+        "skrl_cfg_entry_point": f"{agents.__name__}:skrl_ppo_cfg.yaml",
     },
 )
 
@@ -35,8 +35,8 @@
     disable_env_checker=True,
     kwargs={
         "env_cfg_entry_point": f"{__name__}.allegro_hand_direct_env_cfg:AllegroHandEnvCfg",
-        "rl_games_cfg_entry_point": f"{agents.__name__}:rl_games_direct_ppo_cfg.yaml",
+        "rl_games_cfg_entry_point": f"{agents.__name__}:rl_games_ppo_cfg.yaml",
         "rsl_rl_cfg_entry_point": f"{agents.__name__}.rsl_rl_ppo_cfg:AllegroHandPPORunnerCfg",
-        "skrl_cfg_entry_point": f"{agents.__name__}:skrl_direct_ppo_cfg.yaml",
+        "skrl_cfg_entry_point": f"{agents.__name__}:skrl_ppo_cfg.yaml",
     },
 )