[Task Clean-up][Manager] Dexterous Part 9/11: Add Shadow state and OpenAI manager counterparts

source/isaaclab_tasks/changelog.d/task-cleanup-dex-part09.minor.rst

@@ -0,0 +1,9 @@
+Added
+^^^^^
+
+* Added a manager-based counterpart for the state-based Shadow Hand
+  reorientation environment.
+* Added manager-based counterparts for the Shadow Hand OpenAI feed-forward and
+  recurrent reorientation environments.
+* Added an RSL-RL LSTM training configuration for the Shadow Hand OpenAI
+  reorientation environment.

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,
@@ -48,6 +83,7 @@
     kwargs={
         "env_cfg_entry_point": f"{__name__}.shadow_hand_env_cfg:ShadowHandOpenAIEnvCfg",
         "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",
     },
 )
 

source/isaaclab_tasks/isaaclab_tasks/core/reorient/config/shadow_hand/agents/rsl_rl_ppo_cfg.py

@@ -5,7 +5,7 @@
 
 from isaaclab.utils.configclass import configclass
 
-from isaaclab_rl.rsl_rl import RslRlMLPModelCfg, RslRlOnPolicyRunnerCfg, RslRlPpoAlgorithmCfg
+from isaaclab_rl.rsl_rl import RslRlMLPModelCfg, RslRlOnPolicyRunnerCfg, RslRlPpoAlgorithmCfg, RslRlRNNModelCfg
 
 
 @configclass
@@ -47,6 +47,7 @@ class ShadowHandAsymFFPPORunnerCfg(RslRlOnPolicyRunnerCfg):
     max_iterations = 10000
     save_interval = 250
     experiment_name = "shadow_hand_openai_ff"
+    obs_groups = {"actor": ["policy"], "critic": ["critic"]}
     actor = RslRlMLPModelCfg(
         hidden_dims=[400, 400, 200, 100],
         activation="elu",
@@ -74,12 +75,37 @@ class ShadowHandAsymFFPPORunnerCfg(RslRlOnPolicyRunnerCfg):
     )
 
 
+@configclass
+class ShadowHandAsymLSTMPPORunnerCfg(ShadowHandAsymFFPPORunnerCfg):
+    """RSL-RL recurrent policy configuration for the asymmetric OpenAI observations."""
+
+    experiment_name = "shadow_hand_openai_lstm"
+    actor = RslRlRNNModelCfg(
+        hidden_dims=[400, 400, 200, 100],
+        activation="elu",
+        obs_normalization=True,
+        distribution_cfg=RslRlMLPModelCfg.GaussianDistributionCfg(init_std=1.0),
+        rnn_type="lstm",
+        rnn_hidden_dim=256,
+        rnn_num_layers=1,
+    )
+    critic = RslRlRNNModelCfg(
+        hidden_dims=[512, 512, 256, 128],
+        activation="elu",
+        obs_normalization=True,
+        rnn_type="lstm",
+        rnn_hidden_dim=256,
+        rnn_num_layers=1,
+    )
+
+
 @configclass
 class ShadowHandCameraFFPPORunnerCfg(RslRlOnPolicyRunnerCfg):
     num_steps_per_env = 64
     max_iterations = 50000
     save_interval = 250
     experiment_name = "shadow_hand_camera"
+    obs_groups = {"actor": ["policy"], "critic": ["critic"]}
     actor = RslRlMLPModelCfg(
         hidden_dims=[1024, 512, 512, 256, 128],
         activation="elu",