Add manager-based counterparts for the reorientation tasks

Rebuilt from the reviewed lump so the manager counterparts land with
the review round folded in:

- Sim settings come from the ShadowHandTaskCfgBase/OpenAI and Allegro
  task-cfg mixins instead of module-level cfg variables.
- Backend scene presets declare only backend-specific fields; shared
  invariants live as inner-class defaults.
- The command, reward, action, and reset-event declarations come from
  term factories in reorient_task_base; each variant states only its
  deltas, and the invariant termination section is shared outright.
- The manager terms are named ReorientReward and ReorientTimeout; the
  Direct-parity contract is stated in their docstrings.
- The orientation error comes from isaaclab.utils.math.quat_error_magnitude.
- The superseded legacy manager env and its terms are removed.

Lump review commits folded: 2c22af0, 792e400, 42675b6,
1f05f85, c6140f9, 173e9dc, 2727616, 1732493,
3659a33, ee272c9.

Author: hujc7

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

@@ -0,0 +1,31 @@
+Added
+^^^^^
+
+* Added manager-based counterparts for the Allegro and Shadow cube
+  reorientation tasks (state and OpenAI FF/LSTM observation variants), sharing
+  the Direct tasks' scalar parameters and boolean success metrics through
+  common MDP terms.
+* Added opt-in domain randomization to the manager-based Allegro environment
+  (``enable_domain_randomization``, disabled by default; enabling requires
+  retraining).
+* Added a Newton physics preset to the manager-based Allegro environment.
+
+Removed
+^^^^^^^
+
+* Removed the legacy manager-based reorientation configuration
+  ``ReorientObjectEnvCfg`` and the manager terms only it consumed
+  (``success_bonus``, ``track_pos_l2``, ``track_orientation_inv_l2``,
+  ``max_consecutive_success``, ``object_away_from_goal``, and
+  ``object_away_from_robot``). Use the Direct-compatible manager
+  configurations and terms instead (e.g.
+  :class:`~isaaclab_tasks.core.reorient.mdp.ReorientReward` and
+  :class:`~isaaclab_tasks.core.reorient.mdp.ReorientTimeout`).
+
+Changed
+^^^^^^^
+
+* **Breaking:** Changed the manager-based Allegro reorientation environment to
+  match the Direct observation, action, reward, reset, termination, success,
+  asset, and benchmark contracts. Existing manager checkpoints are
+  incompatible and must be retrained.

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/allegro_hand_manager_env_cfg.py

@@ -3,36 +3,260 @@
 #
 # SPDX-License-Identifier: BSD-3-Clause
 
+"""Manager-based counterpart of the Allegro Hand Direct reorientation task."""
+
+import isaaclab.sim as sim_utils
+from isaaclab.assets import ArticulationCfg, AssetBaseCfg
+from isaaclab.envs import ManagerBasedRLEnvCfg
+from isaaclab.managers import EventTermCfg as EventTerm
+from isaaclab.managers import ObservationGroupCfg as ObsGroup
+from isaaclab.managers import ObservationTermCfg as ObsTerm
+from isaaclab.managers import SceneEntityCfg
+from isaaclab.scene import InteractiveSceneCfg
 from isaaclab.utils.configclass import configclass
 
-from isaaclab_tasks.core.reorient.reorient_manager_env_cfg import ReorientObjectEnvCfg
+import isaaclab_tasks.core.reorient.mdp as mdp
+from isaaclab_tasks.core.reorient.config.allegro_hand.allegro_hand_direct_env_cfg import (
+    GOAL_OBJECT_CFG,
+    OBJECT_CFG,
+    ROBOT_CFG,
+    AllegroHandTaskCfgBase,
+    ObjectCfg,
+)
+from isaaclab_tasks.core.reorient.reorient_task_base import (
+    ALLEGRO_ACTUATED_JOINT_NAMES,
+    ALLEGRO_FINGERTIP_BODY_NAMES,
+    ReorientTerminationsCfg,
+    reorient_goal_command,
+    reorient_joint_action,
+    reorient_reset_event,
+    reorient_reward_term,
+)
+from isaaclab_tasks.utils import PresetCfg
+
+
+@configclass
+class AllegroCubeSceneCfg(PresetCfg):
+    """Backend-specific scene cloning settings matching the Direct task."""
+
+    @configclass
+    class SceneCfg(InteractiveSceneCfg):
+        """Allegro scene shared by the backend alternatives."""
 
-##
-# Pre-defined configs
-##
-from isaaclab_assets import ALLEGRO_HAND_CFG  # isort: skip
+        num_envs = 8192
+        env_spacing = 0.75
+        replicate_physics = True
+
+        ground = AssetBaseCfg(prim_path="/World/ground", spawn=sim_utils.GroundPlaneCfg())
+        robot: ArticulationCfg = ROBOT_CFG
+        object: ObjectCfg = OBJECT_CFG
+        light = AssetBaseCfg(
+            prim_path="/World/Light",
+            spawn=sim_utils.DomeLightCfg(intensity=2000.0, color=(0.75, 0.75, 0.75)),
+        )
+
+    physx = SceneCfg(clone_in_fabric=True)
+    newton_mjwarp = SceneCfg(clone_in_fabric=False)
+    ovphysx = physx
+    default = physx
+
+    def set_num_envs(self, num_envs: int) -> None:
+        """Set the environment count on every backend alternative."""
+        for scene in (self.physx, self.newton_mjwarp, self.ovphysx, self.default):
+            scene.num_envs = num_envs
 
 
 @configclass
-class AllegroCubeEnvCfg(ReorientObjectEnvCfg):
-    def __post_init__(self):
-        # post init of parent
-        super().__post_init__()
+class CommandsCfg:
+    """The reorient goal command with the Allegro marker and tolerance."""
+
+    object_pose = reorient_goal_command(orientation_success_threshold=0.2, goal_pose_visualizer_cfg=GOAL_OBJECT_CFG)
+
+
+@configclass
+class ActionsCfg:
+    """Sixteen actuated Allegro Hand joints in Direct order."""
+
+    joint_pos = reorient_joint_action(ALLEGRO_ACTUATED_JOINT_NAMES)
+
+
+@configclass
+class ObservationsCfg:
+    """Full 124-dimensional state observation in Direct order."""
+
+    @configclass
+    class PolicyCfg(ObsGroup):
+        joint_pos = ObsTerm(
+            func=mdp.joint_pos_limit_normalized,
+            params={"asset_cfg": SceneEntityCfg("robot", joint_names=".*", preserve_order=False)},
+        )
+        joint_vel = ObsTerm(
+            func=mdp.joint_vel,
+            scale=0.2,
+            params={"asset_cfg": SceneEntityCfg("robot", joint_names=".*", preserve_order=False)},
+        )
+        object_pos = ObsTerm(func=mdp.root_pos_w, params={"asset_cfg": SceneEntityCfg("object")})
+        object_quat = ObsTerm(
+            func=mdp.root_quat_w,
+            params={"asset_cfg": SceneEntityCfg("object"), "make_quat_unique": False},
+        )
+        object_lin_vel = ObsTerm(func=mdp.root_lin_vel_w, params={"asset_cfg": SceneEntityCfg("object")})
+        object_ang_vel = ObsTerm(
+            func=mdp.root_ang_vel_w,
+            scale=0.2,
+            params={"asset_cfg": SceneEntityCfg("object")},
+        )
+        goal_pose = ObsTerm(func=mdp.generated_commands, params={"command_name": "object_pose"})
+        goal_quat_diff = ObsTerm(
+            func=mdp.goal_quat_diff,
+            params={"asset_cfg": SceneEntityCfg("object"), "command_name": "object_pose", "make_quat_unique": False},
+        )
+        fingertip_pos = ObsTerm(
+            func=mdp.fingertip_pos,
+            params={
+                "asset_cfg": SceneEntityCfg("robot", body_names=ALLEGRO_FINGERTIP_BODY_NAMES, preserve_order=False)
+            },
+        )
+        fingertip_quat = ObsTerm(
+            func=mdp.fingertip_quat,
+            params={
+                "asset_cfg": SceneEntityCfg("robot", body_names=ALLEGRO_FINGERTIP_BODY_NAMES, preserve_order=False)
+            },
+        )
+        fingertip_vel = ObsTerm(
+            func=mdp.fingertip_vel,
+            params={
+                "asset_cfg": SceneEntityCfg("robot", body_names=ALLEGRO_FINGERTIP_BODY_NAMES, preserve_order=False)
+            },
+        )
+        last_action = ObsTerm(func=mdp.reorient_last_action, params={"action_name": "joint_pos"})
+
+        def __post_init__(self):
+            self.enable_corruption = False
+            self.concatenate_terms = True
 
-        # switch robot to allegro hand
-        self.scene.robot = ALLEGRO_HAND_CFG.replace(prim_path="{ENV_REGEX_NS}/Robot")
-        # enable clone in fabric
-        self.scene.clone_in_fabric = True
+    policy: PolicyCfg = PolicyCfg()
+
+
+@configclass
+class EventCfg:
+    """Reset distributions matching the Direct task, plus opt-in domain randomization.
+
+    The domain-randomization terms reproduce the legacy manager recipe. They are
+    startup-mode terms and are dropped by default (see
+    :attr:`AllegroCubeEnvCfg.enable_domain_randomization`): the Direct task has no
+    domain randomization, and the validated benchmark thresholds were calibrated
+    without it. Enabling them requires retraining.
+    """
+
+    # -- opt-in domain randomization (legacy manager recipe parameters)
+    robot_physics_material = EventTerm(
+        func=mdp.randomize_rigid_body_material,
+        mode="startup",
+        params={
+            "asset_cfg": SceneEntityCfg("robot", body_names=".*"),
+            "static_friction_range": (0.7, 1.3),
+            "dynamic_friction_range": (0.7, 1.3),
+            "restitution_range": (0.0, 0.0),
+            "num_buckets": 250,
+        },
+    )
+    robot_scale_mass = EventTerm(
+        func=mdp.randomize_rigid_body_mass,
+        mode="startup",
+        params={
+            "asset_cfg": SceneEntityCfg("robot", body_names=".*"),
+            "mass_distribution_params": (0.95, 1.05),
+            "operation": "scale",
+        },
+    )
+    robot_joint_stiffness_and_damping = EventTerm(
+        func=mdp.randomize_actuator_gains,
+        mode="startup",
+        params={
+            "asset_cfg": SceneEntityCfg("robot", joint_names=".*"),
+            "stiffness_distribution_params": (0.3, 3.0),
+            "damping_distribution_params": (0.75, 1.5),
+            "operation": "scale",
+            "distribution": "log_uniform",
+        },
+    )
+    object_physics_material = EventTerm(
+        func=mdp.randomize_rigid_body_material,
+        mode="startup",
+        params={
+            "asset_cfg": SceneEntityCfg("object", body_names=".*"),
+            "static_friction_range": (0.7, 1.3),
+            "dynamic_friction_range": (0.7, 1.3),
+            "restitution_range": (0.0, 0.0),
+            "num_buckets": 250,
+        },
+    )
+    object_scale_mass = EventTerm(
+        func=mdp.randomize_rigid_body_mass,
+        mode="startup",
+        params={
+            "asset_cfg": SceneEntityCfg("object"),
+            "mass_distribution_params": (0.4, 1.6),
+            "operation": "scale",
+        },
+    )
+
+    reset_state = reorient_reset_event()
+
+
+@configclass
+class RewardsCfg:
+    """The reorient reward with the Allegro success tolerance."""
+
+    reorient = reorient_reward_term(success_tolerance=0.2)
+
+
+@configclass
+class AllegroCubeEnvCfg(AllegroHandTaskCfgBase, ManagerBasedRLEnvCfg):
+    """Manager-based Allegro Hand task with Direct-compatible semantics."""
+
+    scene: AllegroCubeSceneCfg = AllegroCubeSceneCfg()
+    observations: ObservationsCfg = ObservationsCfg()
+    actions: ActionsCfg = ActionsCfg()
+    commands: CommandsCfg = CommandsCfg()
+    rewards: RewardsCfg = RewardsCfg()
+    # the shared termination section from isaaclab_tasks.core.reorient.reorient_task_base
+    terminations: ReorientTerminationsCfg = ReorientTerminationsCfg()
+    events: EventCfg = EventCfg()
+
+    enable_domain_randomization: bool = False
+    """Enable the legacy startup domain-randomization terms.
+
+    Disabled by default: the validated reference training runs and the benchmark
+    thresholds were produced without domain randomization, so enabling it
+    requires retraining and threshold recalibration.
+    """
+
+    _DOMAIN_RANDOMIZATION_TERMS = (
+        "robot_physics_material",
+        "robot_scale_mass",
+        "robot_joint_stiffness_and_damping",
+        "object_physics_material",
+        "object_scale_mass",
+    )
+
+    def __post_init__(self):
+        self.decimation = 4
+        self.episode_length_s = 10.0
+        self.sim.render_interval = self.decimation
+        self.viewer.eye = (2.0, 2.0, 2.0)
+        if not self.enable_domain_randomization:
+            for term_name in self._DOMAIN_RANDOMIZATION_TERMS:
+                setattr(self.events, term_name, None)
 
 
 @configclass
 class AllegroCubeEnvCfg_PLAY(AllegroCubeEnvCfg):
+    """Reduced, deterministic play configuration."""
+
     def __post_init__(self):
-        # post init of parent
         super().__post_init__()
-        # make a smaller scene for play
-        self.scene.num_envs = 50
-        # disable randomization for play
+        self.scene.set_num_envs(50)
         self.observations.policy.enable_corruption = False
-        # remove termination due to timeouts
         self.terminations.time_out = None

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

@@ -17,6 +17,18 @@
 
 reorient_direct_entry = "isaaclab_tasks.core.reorient.reorient_direct_env:ReorientDirectEnv"
 
+gym.register(
+    id="Isaac-Reorient-Cube-Shadow",
+    entry_point="isaaclab.envs:ManagerBasedRLEnv",
+    disable_env_checker=True,
+    kwargs={
+        "env_cfg_entry_point": f"{__name__}.shadow_hand_manager_env_cfg:ShadowHandManagerEnvCfg",
+        "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:ShadowHandPPORunnerCfg",
+        "skrl_cfg_entry_point": f"{agents.__name__}:skrl_ppo_cfg.yaml",
+    },
+)
+
 gym.register(
     id="Isaac-Reorient-Cube-Shadow-Direct",
     entry_point=reorient_direct_entry,
@@ -29,6 +41,29 @@
     },
 )
 
+gym.register(
+    id="Isaac-Reorient-Cube-Shadow-OpenAI-FF",
+    entry_point="isaaclab.envs:ManagerBasedRLEnv",
+    disable_env_checker=True,
+    kwargs={
+        "env_cfg_entry_point": f"{__name__}.shadow_hand_manager_env_cfg:ShadowHandOpenAIManagerEnvCfg",
+        "rl_games_cfg_entry_point": f"{agents.__name__}:rl_games_ppo_ff_cfg.yaml",
+        "rsl_rl_cfg_entry_point": f"{agents.__name__}.rsl_rl_ppo_cfg:ShadowHandAsymFFPPORunnerCfg",
+        "skrl_cfg_entry_point": f"{agents.__name__}:skrl_ff_ppo_cfg.yaml",
+    },
+)
+
+gym.register(
+    id="Isaac-Reorient-Cube-Shadow-OpenAI-LSTM",
+    entry_point="isaaclab.envs:ManagerBasedRLEnv",
+    disable_env_checker=True,
+    kwargs={
+        "env_cfg_entry_point": f"{__name__}.shadow_hand_manager_env_cfg:ShadowHandOpenAIManagerEnvCfg",
+        "rl_games_cfg_entry_point": f"{agents.__name__}:rl_games_ppo_lstm_cfg.yaml",
+        "rsl_rl_cfg_entry_point": f"{agents.__name__}.rsl_rl_ppo_cfg:ShadowHandAsymLSTMPPORunnerCfg",
+    },
+)
+
 gym.register(
     id="Isaac-Reorient-Cube-Shadow-OpenAI-FF-Direct",
     entry_point=reorient_direct_entry,