Fix the bug in D4RL that the seed parameter is ignored.

d4rl/locomotion/ant.py

@@ -204,8 +204,8 @@ def reset(self):
   def set_target(self, target_location=None):
     return self.set_target_goal(target_location)
 
-  def seed(self, seed=0):
-      mujoco_env.MujocoEnv.seed(self, seed)
+  def seed(self, seed=None):
+    mujoco_env.MujocoEnv.seed(self, seed)
 
 def make_ant_maze_env(**kwargs):
   env = AntMazeEnv(**kwargs)

d4rl/locomotion/maze_env.py

@@ -214,13 +214,13 @@ def _xy_to_rowcol(self, xy):
   def _get_reset_location(self,):
     prob = (1.0 - self._np_maze_map) / np.sum(1.0 - self._np_maze_map) 
     prob_row = np.sum(prob, 1)
-    row_sample = np.random.choice(np.arange(self._np_maze_map.shape[0]), p=prob_row)
-    col_sample = np.random.choice(np.arange(self._np_maze_map.shape[1]), p=prob[row_sample] * 1.0 / prob_row[row_sample])
+    row_sample = self.np_random.choice(np.arange(self._np_maze_map.shape[0]), p=prob_row)
+    col_sample = self.np_random.choice(np.arange(self._np_maze_map.shape[1]), p=prob[row_sample] * 1.0 / prob_row[row_sample])
     reset_location = self._rowcol_to_xy((row_sample, col_sample))
 
     # Add some random noise
-    random_x = np.random.uniform(low=0, high=0.5) * 0.5 * self._maze_size_scaling
-    random_y = np.random.uniform(low=0, high=0.5) * 0.5 * self._maze_size_scaling
+    random_x = self.np_random.uniform(low=0, high=0.5) * 0.5 * self._maze_size_scaling
+    random_y = self.np_random.uniform(low=0, high=0.5) * 0.5 * self._maze_size_scaling
 
     return (max(reset_location[0] + random_x, 0), max(reset_location[1] + random_y, 0))
 
@@ -229,8 +229,8 @@ def _rowcol_to_xy(self, rowcol, add_random_noise=False):
     x = col * self._maze_size_scaling - self._init_torso_x
     y = row * self._maze_size_scaling - self._init_torso_y
     if add_random_noise:
-      x = x + np.random.uniform(low=0, high=self._maze_size_scaling * 0.25)
-      y = y + np.random.uniform(low=0, high=self._maze_size_scaling * 0.25)
+      x = x + self.np_random.uniform(low=0, high=self._maze_size_scaling * 0.25)
+      y = y + self.np_random.uniform(low=0, high=self._maze_size_scaling * 0.25)
     return (x, y)
 
   def goal_sampler(self, np_random, only_free_cells=True, interpolate=True):
@@ -247,19 +247,19 @@ def goal_sampler(self, np_random, only_free_cells=True, interpolate=True):
     # If there is a 'goal' designated, use that. Otherwise, any valid cell can
     # be a goal.
     sample_choices = goal_cells if goal_cells else valid_cells
-    cell = sample_choices[np_random.choice(len(sample_choices))]
+    cell = sample_choices[self.np_random.choice(len(sample_choices))]
     xy = self._rowcol_to_xy(cell, add_random_noise=True)
 
-    random_x = np.random.uniform(low=0, high=0.5) * 0.25 * self._maze_size_scaling
-    random_y = np.random.uniform(low=0, high=0.5) * 0.25 * self._maze_size_scaling
+    random_x = self.np_random.uniform(low=0, high=0.5) * 0.25 * self._maze_size_scaling
+    random_y = self.np_random.uniform(low=0, high=0.5) * 0.25 * self._maze_size_scaling
 
     xy = (max(xy[0] + random_x, 0), max(xy[1] + random_y, 0))
 
     return xy
 
   def set_target_goal(self, goal_input=None):
     if goal_input is None:
-      self.target_goal = self.goal_sampler(np.random)
+      self.target_goal = self.goal_sampler(self.np_random)
     else:
       self.target_goal = goal_input
 

d4rl/locomotion/wrappers.py

@@ -22,6 +22,9 @@ def reset(self, **kwargs):
 
     def step(self, action):
         return self._wrapped_env.step(action)
+
+    def seed(self, seed=None):
+        return self._wrapped_env.seed(seed)
 
     def render(self, *args, **kwargs):
         return self._wrapped_env.render(*args, **kwargs)