Fixes device decoupling of RL Games training vs. sim device

scripts/reinforcement_learning/rl_games/play.py

@@ -95,10 +95,6 @@ def main(env_cfg: ManagerBasedRLEnvCfg | DirectRLEnvCfg | DirectMARLEnvCfg, agen
     # override configurations with non-hydra CLI arguments
     env_cfg.scene.num_envs = args_cli.num_envs if args_cli.num_envs is not None else env_cfg.scene.num_envs
     env_cfg.sim.device = args_cli.device if args_cli.device is not None else env_cfg.sim.device
-    # update agent device to match simulation device
-    if args_cli.device is not None:
-        agent_cfg["params"]["config"]["device"] = args_cli.device
-        agent_cfg["params"]["config"]["device_name"] = args_cli.device
 
     # randomly sample a seed if seed = -1
     if args_cli.seed == -1:

scripts/reinforcement_learning/rl_games/train.py

@@ -102,11 +102,6 @@ def main(env_cfg: ManagerBasedRLEnvCfg | DirectRLEnvCfg | DirectMARLEnvCfg, agen
             "Please use GPU device (e.g., --device cuda) for distributed training."
         )
 
-    # update agent device to match simulation device
-    if args_cli.device is not None:
-        agent_cfg["params"]["config"]["device"] = args_cli.device
-        agent_cfg["params"]["config"]["device_name"] = args_cli.device
-
     # randomly sample a seed if seed = -1
     if args_cli.seed == -1:
         args_cli.seed = random.randint(0, 10000)

source/isaaclab_rl/config/extension.toml

@@ -1,7 +1,7 @@
 [package]
 
 # Note: Semantic Versioning is used: https://semver.org/
-version = "0.4.4"
+version = "0.4.5"
 
 # Description
 title = "Isaac Lab RL"

source/isaaclab_rl/docs/CHANGELOG.rst

@@ -1,6 +1,16 @@
 Changelog
 ---------
 
+0.4.5 (2025-11-10)
+~~~~~~~~~~~~~~~~~~
+
+Changed
+^^^^^^^
+
+* Added support for decoupling RL device from simulation device in for RL games wrapper.
+  This allows users to run simulation on one device (e.g., CPU) while running RL training/inference on another device.
+
+
 0.4.4 (2025-10-15)
 ~~~~~~~~~~~~~~~~~~
 

source/isaaclab_rl/isaaclab_rl/rl_games/rl_games.py

@@ -319,6 +319,10 @@ def _process_obs(self, obs_dict: VecEnvObs) -> dict[str, torch.Tensor] | dict[st
             - ``"obs"``: either a concatenated tensor (``concate_obs_group=True``) or a Dict of group tensors.
             - ``"states"`` (optional): same structure as above when state groups are configured; omitted otherwise.
         """
+        # move observations to RL device if different from sim device
+        if self._rl_device != self._sim_device:
+            obs_dict = {key: obs.to(device=self._rl_device) for key, obs in obs_dict.items()}
+
         # clip the observations
         for key, obs in obs_dict.items():
             obs_dict[key] = torch.clamp(obs, -self._clip_obs, self._clip_obs)