Fix zero-life reset error in Atari env (#175)

add `info["terminated"]` as an indicator of `env.game_over()`

envpool/__init__.py

@@ -23,7 +23,7 @@
   register,
 )
 
-__version__ = "0.6.2.post2"
+__version__ = "0.6.3"
 __all__ = [
   "register",
   "make",

envpool/atari/api_test.py

@@ -259,7 +259,7 @@ def test_lowlevel_step(self) -> None:
     np.testing.assert_allclose(done.shape, (num_envs,))
     self.assertEqual(done.dtype, np.bool_)
     self.assertIsInstance(info, dict)
-    self.assertEqual(len(info), 5)
+    self.assertEqual(len(info), 6)
     self.assertEqual(info["env_id"].dtype, np.int32)
     self.assertEqual(info["lives"].dtype, np.int32)
     self.assertEqual(info["players"]["env_id"].dtype, np.int32)
@@ -295,7 +295,7 @@ def test_highlevel_step(self) -> None:
     np.testing.assert_allclose(done.shape, (num_envs,))
     self.assertEqual(done.dtype, np.bool_)
     self.assertIsInstance(info, dict)
-    self.assertEqual(len(info), 5)
+    self.assertEqual(len(info), 6)
     self.assertEqual(info["env_id"].dtype, np.int32)
     self.assertEqual(info["lives"].dtype, np.int32)
     self.assertEqual(info["players"]["env_id"].dtype, np.int32)

envpool/atari/atari_env.h

@@ -65,8 +65,9 @@ class AtariEnvFns {
                          conf["img_height"_], conf["img_width"_]},
                         {0, 255})),
                     "discount"_.Bind(Spec<float>({-1}, {0.0, 1.0})),
-                    "info:lives"_.Bind(Spec<int>({-1}, {0, 5})),
-                    "info:reward"_.Bind(Spec<float>({-1})));
+                    "info:lives"_.Bind(Spec<int>({-1})),
+                    "info:reward"_.Bind(Spec<float>({-1})),
+                    "info:terminated"_.Bind(Spec<int>({-1}, {0, 1})));
   }
   template <typename Config>
   static decltype(auto) ActionSpec(const Config& conf) {
@@ -199,7 +200,7 @@ class AtariEnv : public Env<AtariEnvSpec> {
     PushStack(false, skip_id == 0);
     ++elapsed_step_;
     done_ |= (elapsed_step_ >= max_episode_steps_);
-    if (episodic_life_ && env_->lives() < lives_) {
+    if (episodic_life_ && 0 < env_->lives() && env_->lives() < lives_) {
       done_ = true;
     }
     float discount;
@@ -229,6 +230,7 @@ class AtariEnv : public Env<AtariEnvSpec> {
     state["reward"_] = reward;
     state["info:lives"_] = lives_;
     state["info:reward"_] = info_reward;
+    state["info:terminated"_] = env_->game_over();
     for (int i = 0; i < stack_num_; ++i) {
       state["obs"_]
           .Slice(gray_scale_ ? i : i * 3, gray_scale_ ? i + 1 : (i + 1) * 3)

envpool/atari/atari_envpool_test.py

@@ -78,6 +78,51 @@ def test_align(self) -> None:
       np.testing.assert_allclose(obs0, obs1)
       # cv2.imwrite(f"/tmp/log/align{i}.png", obs0[0, 1:].transpose(1, 2, 0))
 
+  def test_reset_life(self) -> None:
+    """Issue 171."""
+    for env_id in [
+      "atlantis", "backgammon", "breakout", "pong", "wizard_of_wor"
+    ]:
+      np.random.seed(0)
+      env = AtariGymEnvPool(
+        AtariEnvSpec(
+          AtariEnvSpec.gen_config(task=env_id, num_envs=1, episodic_life=True)
+        )
+      )
+      action_num = env.action_space.n  # type: ignore
+      env.reset()
+      info = env.step(np.array([0]))[-1]
+      if info["lives"].sum() == 0:
+        # no life in this game
+        continue
+      for _ in range(10000):
+        _, _, done, info = env.step(np.random.randint(0, action_num, 1))
+        if info["lives"][0] == 0:
+          break
+        else:
+          self.assertFalse(info["terminated"][0])
+      if info["lives"][0] > 0:
+        # step too long
+        continue
+      # for normal atari (e.g., breakout)
+      # take an additional step after all lives are exhausted
+      _, _, next_done, next_info = env.step(
+        np.random.randint(0, action_num, 1)
+      )
+      if done[0] and next_info["lives"][0] > 0:
+        self.assertTrue(info["terminated"][0])
+        continue
+      self.assertFalse(done[0])
+      self.assertFalse(info["terminated"][0])
+      while not done[0]:
+        self.assertFalse(info["terminated"][0])
+        _, _, done, info = env.step(np.random.randint(0, action_num, 1))
+      _, _, next_done, next_info = env.step(
+        np.random.randint(0, action_num, 1)
+      )
+      self.assertTrue(next_info["lives"][0] > 0)
+      self.assertTrue(info["terminated"][0])
+
   def test_partial_step(self) -> None:
     num_envs = 5
     max_episode_steps = 10

setup.cfg

@@ -1,6 +1,6 @@
 [metadata]
 name = envpool
-version = 0.6.2.post2
+version = 0.6.3
 author = "EnvPool Contributors"
 author_email = "sail@sea.com"
 description = "C++-based high-performance parallel environment execution engine (vectorized env) for general RL environments."