Remove RLFramework enum, replace with plain string (#544)

## Summary
- Deletes `isaaclab_arena/reinforcement_learning/`: A module that
existed solely for an `RLFramework` enum whose only method produced
strings like `"rsl_rl_cfg_entry_point"`
- Replaces `rl_framework: RLFramework` with `rl_framework_entry_point:
str` on `IsaacLabArenaEnvironment` and let the user pass in the correct
string directly

Signed-off-by: Clemens Volk <cvolk@nvidia.com>

Author: cvolkcvolk

docs/pages/example_workflows/dexsuite_lift/step_1_environment_setup.rst

@@ -32,7 +32,6 @@ The environment is defined in
               import isaaclab_tasks.manager_based.manipulation.dexsuite  # noqa: F401
 
               from isaaclab_arena.environments.isaaclab_arena_environment import IsaacLabArenaEnvironment
-              from isaaclab_arena.reinforcement_learning.frameworks import RLFramework
               from isaaclab_arena.scene.scene import Scene
               from isaaclab_arena.tasks.lift_object_task import DexsuiteLiftTask
               from isaaclab_arena.utils.pose import Pose, PoseRange
@@ -69,7 +68,7 @@ The environment is defined in
                   scene=scene,
                   task=task,
                   teleop_device=None,
-                  rl_framework=RLFramework.RSL_RL,
+                  rl_framework_entry_point="rsl_rl_cfg_entry_point",
                   rl_policy_cfg=dexsuite_rl_cfg_entry,
               )
 

isaaclab_arena/environments/arena_env_builder.py

@@ -315,9 +315,8 @@ def build_registered(
         kwargs = {
             "env_cfg_entry_point": cfg_entry,
         }
-        if self.arena_env.rl_framework is not None:
-            assert self.arena_env.rl_policy_cfg is not None
-            kwargs[self.arena_env.rl_framework.get_entry_point_string()] = self.arena_env.rl_policy_cfg
+        if self.arena_env.rl_framework_entry_point is not None:
+            kwargs[self.arena_env.rl_framework_entry_point] = self.arena_env.rl_policy_cfg
         gym.register(
             id=name,
             entry_point=entry_point,

isaaclab_arena/environments/isaaclab_arena_environment.py

@@ -13,7 +13,6 @@
     from isaaclab_arena.embodiments.embodiment_base import EmbodimentBase
     from isaaclab_arena.environments.isaaclab_arena_manager_based_env import IsaacLabArenaManagerBasedRLEnvCfg
     from isaaclab_arena.orchestrator.orchestrator_base import OrchestratorBase
-    from isaaclab_arena.reinforcement_learning.frameworks import RLFramework
     from isaaclab_arena.scene.scene import Scene
     from isaaclab_arena.tasks.task_base import TaskBase
 
@@ -30,7 +29,7 @@ def __init__(
         teleop_device: TeleopDeviceBase | None = None,
         orchestrator: OrchestratorBase | None = None,
         env_cfg_callback: Callable[IsaacLabArenaManagerBasedRLEnvCfg] | None = None,
-        rl_framework: RLFramework | None = None,
+        rl_framework_entry_point: str | None = None,
         rl_policy_cfg: str | None = None,
     ):
         """
@@ -42,6 +41,14 @@ def __init__(
             teleop_device: The teleop device to use in the environment.
             orchestrator: The orchestrator to use in the environment.
             env_cfg_callback: A callback function that modifies the environment configuration.
+            rl_framework_entry_point: Gym kwargs key under which the RL policy config is
+                registered. This is an IsaacLab convention: each supported RL framework has a
+                fixed key that its training scripts look up via ``load_cfg_from_registry``.
+                Common values: ``"rsl_rl_cfg_entry_point"``, ``"skrl_cfg_entry_point"``,
+                ``"sb3_cfg_entry_point"``, ``"rl_games_cfg_entry_point"``. Required when
+                ``rl_policy_cfg`` is set.
+            rl_policy_cfg: Import path to the RL policy config class, e.g.
+                ``"my_module:RLPolicyCfg"``.
         """
         self.name = name
         self.scene = scene
@@ -50,5 +57,7 @@ def __init__(
         self.teleop_device = teleop_device
         self.orchestrator = orchestrator
         self.env_cfg_callback = env_cfg_callback
-        self.rl_framework = rl_framework
+        if (rl_framework_entry_point is None) != (rl_policy_cfg is None):
+            raise ValueError("rl_framework_entry_point and rl_policy_cfg must both be set or both be None.")
+        self.rl_framework_entry_point = rl_framework_entry_point
         self.rl_policy_cfg = rl_policy_cfg

isaaclab_arena/reinforcement_learning/frameworks.py

@@ -27,7 +27,6 @@ def get_env(self, args_cli: argparse.Namespace):
         import isaaclab_tasks.manager_based.manipulation.dexsuite  # noqa: F401
 
         from isaaclab_arena.environments.isaaclab_arena_environment import IsaacLabArenaEnvironment
-        from isaaclab_arena.reinforcement_learning.frameworks import RLFramework
         from isaaclab_arena.scene.scene import Scene
         from isaaclab_arena.tasks.lift_object_task import DexsuiteLiftTask
         from isaaclab_arena.utils.pose import Pose, PoseRange
@@ -64,7 +63,7 @@ def get_env(self, args_cli: argparse.Namespace):
             scene=scene,
             task=task,
             teleop_device=None,
-            rl_framework=RLFramework.RSL_RL,
+            rl_framework_entry_point="rsl_rl_cfg_entry_point",
             rl_policy_cfg=dexsuite_rl_cfg_entry,
         )
 

isaaclab_arena_environments/dexsuite_lift_environment.py

@@ -21,7 +21,6 @@ class LiftObjectEnvironment(ExampleEnvironmentBase):
     def get_env(self, args_cli: argparse.Namespace):  # -> IsaacLabArenaEnvironment:
         import isaaclab_arena_examples.policy.base_rsl_rl_policy as base_rsl_rl_policy
         from isaaclab_arena.environments.isaaclab_arena_environment import IsaacLabArenaEnvironment
-        from isaaclab_arena.reinforcement_learning.frameworks import RLFramework
         from isaaclab_arena.scene.scene import Scene
         from isaaclab_arena.tasks.lift_object_task import LiftObjectTaskRL
         from isaaclab_arena.utils.pose import Pose
@@ -66,7 +65,7 @@ def get_env(self, args_cli: argparse.Namespace):  # -> IsaacLabArenaEnvironment:
             scene=scene,
             task=task,
             teleop_device=teleop_device,
-            rl_framework=RLFramework.RSL_RL,
+            rl_framework_entry_point="rsl_rl_cfg_entry_point",
             rl_policy_cfg=f"{base_rsl_rl_policy.__name__}:RLPolicyCfg",
         )