added the option for a deterministic run

envs/crafter.py

@@ -5,7 +5,7 @@ import numpy as np
 class Crafter:
     metadata = {}
 
-    def __init__(self, task, size=(64, 64), seed=None):
+    def __init__(self, task, size=(64, 64), seed=0):
         assert task in ("reward", "noreward")
         import crafter
 

envs/dmc.py

@@ -5,14 +5,18 @@ import numpy as np
 class DeepMindControl:
     metadata = {}
 
-    def __init__(self, name, action_repeat=1, size=(64, 64), camera=None):
+    def __init__(self, name, action_repeat=1, size=(64, 64), camera=None, seed=0):
         domain, task = name.split("_", 1)
         if domain == "cup":  # Only domain with multiple words.
             domain = "ball_in_cup"
         if isinstance(domain, str):
             from dm_control import suite
 
-            self._env = suite.load(domain, task)
+            self._env = suite.load(
+                domain,
+                task,
+                task_kwargs={"random": seed},
+            )
         else:
             assert task is None
             self._env = domain()

envs/memorymaze.py

@@ -5,11 +5,11 @@ import numpy as np
 
 
 class MemoryMaze:
-    def __init__(self, task, obs_key="image", act_key="action", size=(64, 64)):
+    def __init__(self, task, obs_key="image", act_key="action", size=(64, 64), seed=0):
         if task == "9x9":
-            self._env = gym.make("memory_maze:MemoryMaze-9x9-v0")
+            self._env = gym.make("memory_maze:MemoryMaze-9x9-v0", seed=seed)
         elif task == "15x15":
-            self._env = gym.make("memory_maze:MemoryMaze-15x15-v0")
+            self._env = gym.make("memory_maze:MemoryMaze-15x15-v0", seed=seed)
         else:
             raise NotImplementedError(task)
         self._obs_is_dict = hasattr(self._env.observation_space, "spaces")