diff --git a/scrapy/downloadermiddlewares/robotstxt.py b/scrapy/downloadermiddlewares/robotstxt.py
index c061c2407..7f6f0d012 100644
--- a/scrapy/downloadermiddlewares/robotstxt.py
+++ b/scrapy/downloadermiddlewares/robotstxt.py
@@ -8,7 +8,9 @@ import logging
 
 from six.moves.urllib import robotparser
 
+from twisted.internet import reactor
 from twisted.internet.defer import Deferred, maybeDeferred
+from twisted.internet.task import deferLater
 from scrapy.exceptions import NotConfigured, IgnoreRequest
 from scrapy.http import Request
 from scrapy.utils.httpobj import urlparse_cached
@@ -57,7 +59,13 @@ class RobotsTxtMiddleware(object):
                 priority=self.DOWNLOAD_PRIORITY,
                 meta={'dont_obey_robotstxt': True}
             )
-            dfd = self.crawler.engine.download(robotsreq, spider)
+            # engine.download() can return an already-called deferred, e.g. if a
+            # middleware returns a response in process_request(). Using
+            # deferLater() ensures that the error callback isn't called
+            # immediately upon being added, so that it doesn't remove the key
+            # before we check for it.
+            dfd = deferLater(reactor, 0, self.crawler.engine.download,
+                             robotsreq, spider)
             dfd.addCallback(self._parse_robots, netloc)
             dfd.addErrback(self._logerror, robotsreq, spider)
             dfd.addErrback(self._robots_error, netloc)