Add depth rendering from camera (openai#1114)

* Add depth rendering from camera

Allow the mujoco renderer to return the depth image using the `depth_array` mode. 

It follows the `rgb_array` extraction method.

* Depth data is a 2d array.

* Add depth_array to render modes.

gym/envs/mujoco/mujoco_env.py

@@ -33,7 +33,7 @@ def __init__(self, model_path, frame_skip):
         self._viewers = {}
 
         self.metadata = {
-            'render.modes': ['human', 'rgb_array'],
+            'render.modes': ['human', 'rgb_array', 'depth_array'],
             'video.frames_per_second': int(np.round(1.0 / self.dt))
         }
 
@@ -112,6 +112,13 @@ def render(self, mode='human', width=DEFAULT_SIZE, height=DEFAULT_SIZE):
             data = self._get_viewer(mode).read_pixels(width, height, depth=False)
             # original image is upside-down, so flip it
             return data[::-1, :, :]
+        elif mode == 'depth_array':
+            self._get_viewer(mode).render(width, height)
+            # window size used for old mujoco-py:
+            # Extract depth part of the read_pixels() tuple
+            data = self._get_viewer(mode).read_pixels(width, height, depth=True)[1]
+            # original image is upside-down, so flip it
+            return data[::-1, :]
         elif mode == 'human':
             self._get_viewer(mode).render()
 
@@ -126,8 +133,10 @@ def _get_viewer(self, mode):
         if self.viewer is None:
             if mode == 'human':
                 self.viewer = mujoco_py.MjViewer(self.sim)
-            elif mode == 'rgb_array':
+
+            elif mode == 'rgb_array' or mode == 'depth_array':
                 self.viewer = mujoco_py.MjRenderContextOffscreen(self.sim, -1)
+
             self.viewer_setup()
             self._viewers[mode] = self.viewer
         return self.viewer