Simplify the TorchRL PPO runner

Let ClipPPOLoss own the GAE estimator, sum whichever loss_* terms the
loss returns instead of naming them, build the networks in two small
module-level helpers, and collapse logging into one dictionary.
Activations are named by their torch.nn class so no lookup table is
needed.

Author: theap06

source/isaaclab_rl/isaaclab_rl/torchrl/ppo_cfg.py

@@ -12,86 +12,53 @@
 
 @configclass
 class TorchRlMlpModelCfg:
-    """Configuration of an MLP used by :class:`~isaaclab_rl.torchrl.TorchRlPpoRunner`."""
+    """MLP used for the actor and the critic of :class:`~isaaclab_rl.torchrl.TorchRlPpoRunner`."""
 
     hidden_dims: list[int] = MISSING
     """Sizes of the hidden layers."""
 
-    activation: str = "elu"
-    """Activation function, one of ``"elu"``, ``"relu"``, ``"tanh"``, ``"selu"``, ``"gelu"`` or ``"silu"``."""
+    activation: str = "ELU"
+    """Name of a :mod:`torch.nn` activation class, e.g. ``"ELU"``, ``"ReLU"`` or ``"Tanh"``."""
 
 
 @configclass
 class TorchRlPpoAlgorithmCfg:
-    """Configuration of the PPO update, mapped onto :class:`torchrl.objectives.ClipPPOLoss`."""
+    """PPO hyper-parameters, mapped onto :class:`torchrl.objectives.ClipPPOLoss` and GAE."""
 
     num_learning_epochs: int = MISSING
-    """Number of passes over the collected batch per iteration."""
-
     num_mini_batches: int = MISSING
-    """Number of mini-batches the collected batch is split into per epoch."""
-
     learning_rate: float = MISSING
-    """Adam learning rate."""
-
     gamma: float = MISSING
-    """Discount factor."""
-
     lam: float = MISSING
-    """GAE lambda."""
-
     clip_param: float = 0.2
-    """PPO clipping range for the probability ratio."""
-
     entropy_coef: float = 0.0
-    """Weight of the entropy bonus. ``0`` disables it."""
-
+    """Weight of the entropy bonus; ``0`` disables it."""
     value_loss_coef: float = 1.0
-    """Weight of the value loss."""
-
     max_grad_norm: float = 1.0
-    """Gradient norm clipping threshold."""
-
     normalize_advantage: bool = True
-    """Whether to normalize advantages within each mini-batch."""
+    """Normalize advantages within each mini-batch."""
 
 
 @configclass
 class TorchRlPpoRunnerCfg:
     """Configuration of :class:`~isaaclab_rl.torchrl.TorchRlPpoRunner`."""
 
     seed: int = 42
-    """Seed for the environment and the networks."""
-
     device: str = "cuda:0"
-    """Device the networks and the collected batches live on."""
-
     num_steps_per_env: int = MISSING
     """Environment steps collected from every environment per iteration."""
-
     max_iterations: int = MISSING
-    """Number of training iterations."""
-
     save_interval: int = MISSING
     """Iterations between checkpoints."""
-
     experiment_name: str = MISSING
     """Name of the experiment folder under ``logs/torchrl``."""
-
     run_name: str = ""
-    """Optional suffix appended to the timestamped run folder."""
-
+    """Optional suffix of the timestamped run folder."""
     clip_actions: float | None = None
-    """Clipping range applied to actions before they reach the environment. ``None`` disables clipping."""
-
+    """Clipping range applied to actions before they reach the environment; ``None`` disables it."""
     init_noise_std: float = 1.0
     """Initial standard deviation of the Gaussian policy."""
-
     actor: TorchRlMlpModelCfg = MISSING
-    """Actor network."""
-
     critic: TorchRlMlpModelCfg = MISSING
-    """Critic network. It reads the ``"critic"`` observation group when the task defines one, else ``"policy"``."""
-
+    """Reads the ``"critic"`` observation group when the task defines one, else ``"policy"``."""
     algorithm: TorchRlPpoAlgorithmCfg = MISSING
-    """PPO hyper-parameters."""