Add action result and alerts observations.

PiperOrigin-RevId: 196657275

docs/environment.md

@@ -322,6 +322,16 @@ observation.
 A `(n)` tensor listing all the action ids that were made successfully since the
 last observation. An action that was attempted but failed is not included here.
 
+##### Action Result
+
+A `(n)` tensor (usually size 1) giving the result of the action. The values are
+listed in
+[error.proto](https://github.com/Blizzard/s2client-proto/blob/master/s2clientprotocol/error.proto)
+
+##### Alerts
+
+A `(n)` tensor (usually empty, occasionally size 1, max 2) for when you're being attacked in a major way.
+
 ### Actions
 
 The SC2 action space is very big. There are hundreds of possible actions, many

pysc2/lib/features.py

@@ -499,6 +499,8 @@ def observation_spec(self):
       units you have selected.
     """
     obs_spec = named_array.NamedDict({
+        "action_result": (0,),  # See error.proto: ActionResult.
+        "alerts": (0,),  # See sc2api.proto: Alert.
         "available_actions": (0,),
         "build_queue": (0, len(UnitLayer)),  # pytype: disable=wrong-arg-types
         "cargo": (0, len(UnitLayer)),  # pytype: disable=wrong-arg-types
@@ -574,6 +576,11 @@ def or_zeros(layer, size):
         [self.reverse_action(a).function for a in obs.actions],
         dtype=np.int32)
 
+    out["action_result"] = np.array([o.result for o in obs.action_errors],
+                                    dtype=np.int32)
+
+    out["alerts"] = np.array(obs.observation.alerts, dtype=np.int32)
+
     out["game_loop"] = np.array([obs.observation.game_loop], dtype=np.int32)
 
     score_details = obs.observation.score.score_details