fix the way we detect np.float64 observations

Summary: The current way to detect np array of type float64 is incorrect. np.float64 is the type of a single number, not the type of an array.

Reviewed By: rodrigodesalvobraz

Differential Revision: D67272396

fbshipit-source-id: 00b4255e65cb65c776a50e65e8f341ad18cbb64d

pearl/utils/instantiations/environments/gym_environment.py

@@ -112,7 +112,7 @@ def reset(self, seed: int | None = None) -> tuple[Observation, ActionSpace]:
             # TODO: Deprecate this part at some point and only support new
             # version of Gymnasium?
             observation = list(reset_result.values())[0]  # pyre-ignore
-        if isinstance(observation, np.float64):
+        if isinstance(observation, np.ndarray) and observation.dtype == np.float64:
             observation = observation.astype(np.float32)
         return observation, self.action_space
 
@@ -151,7 +151,7 @@ def step(self, action: Action) -> ActionResult:
         else:
             available_action_space = None
 
-        if isinstance(observation, np.float64):
+        if isinstance(observation, np.ndarray) and observation.dtype == np.float64:
             observation = observation.astype(np.float32)
         if isinstance(reward, np.float64):
             reward = reward.astype(np.float32)