diff --git a/agent/DQN_agent.py b/agent/DQN_agent.py
index 199604b..80c3a16 100644
--- a/agent/DQN_agent.py
+++ b/agent/DQN_agent.py
@@ -46,10 +46,11 @@ class DQNAgent:
             self.history_buffer.pop(0)
             self.history_buffer.append(next_state)
             next_state = np.vstack(self.history_buffer)
+            total_reward += np.sum(reward * self.config.reward_weight)
+            reward = self.config.reward_shift_fn(reward)
             if not deterministic:
                 self.replay.feed([state, action, reward, next_state, int(done)])
                 self.total_steps += 1
-            total_reward += np.sum(reward * self.config.reward_weight)
             steps += 1
             state = next_state
             if done:
diff --git a/async_worker/actor_critic.py b/async_worker/actor_critic.py
index 9ad9417..e821da6 100644
--- a/async_worker/actor_critic.py
+++ b/async_worker/actor_critic.py
@@ -33,6 +33,7 @@ class AdvantageActorCritic:
 
             steps += 1
             total_reward += reward
+            reward = config.reward_shift_fn(reward)
 
             if deterministic:
                 if terminal:
diff --git a/async_worker/n_step_q.py b/async_worker/n_step_q.py
index 3a5899b..47c4cfb 100644
--- a/async_worker/n_step_q.py
+++ b/async_worker/n_step_q.py
@@ -34,6 +34,7 @@ class NStepQLearning:
 
             steps += 1
             total_reward += reward
+            reward = config.reward_shift_fn(reward)
 
             if deterministic:
                 if terminal:
diff --git a/async_worker/one_step_q.py b/async_worker/one_step_q.py
index d190753..a37bc4c 100644
--- a/async_worker/one_step_q.py
+++ b/async_worker/one_step_q.py
@@ -34,6 +34,7 @@ class OneStepQLearning:
 
             steps += 1
             total_reward += reward
+            reward = config.reward_shift_fn(reward)
 
             if deterministic:
                 if terminal:
diff --git a/async_worker/one_step_sarsa.py b/async_worker/one_step_sarsa.py
index d763867..1434108 100644
--- a/async_worker/one_step_sarsa.py
+++ b/async_worker/one_step_sarsa.py
@@ -37,6 +37,7 @@ class OneStepSarsa:
 
             steps += 1
             total_reward += reward
+            reward = config.reward_shift_fn(reward)
 
             if deterministic:
                 if terminal:
diff --git a/component/atari_wrapper.py b/component/atari_wrapper.py
index ac16b9e..acd8e90 100644
--- a/component/atari_wrapper.py
+++ b/component/atari_wrapper.py
@@ -176,11 +176,6 @@ class ProcessFrame(gym.Wrapper):
     def _reset(self):
         return self.process_fn(self.env.reset())
 
-class ClippedRewardsWrapper(gym.Wrapper):
-    def _step(self, action):
-        obs, reward, done, info = self.env.step(action)
-        return obs, np.sign(reward), done, info
-
 class NormalizeFrame(gym.Wrapper):
     def __init__(self, env=None):
         super(NormalizeFrame, self).__init__(env)
diff --git a/component/task.py b/component/task.py
index d52d009..8079d34 100644
--- a/component/task.py
+++ b/component/task.py
@@ -70,8 +70,7 @@ class PixelAtari(BasicTask):
         env = MaxAndSkipEnv(env, skip=frame_skip)
         if 'FIRE' in env.unwrapped.get_action_meanings():
             env = FireResetEnv(env)
-        env = ProcessFrame(env, frame_size)
-        self.env = ClippedRewardsWrapper(env)
+        self.env = ProcessFrame(env, frame_size)
         self.action_dim = self.env.action_space.n
 
     def normalize_state(self, state):
diff --git a/main.py b/main.py
index 9b74e82..26ba78c 100644
--- a/main.py
+++ b/main.py
@@ -79,6 +79,7 @@ def dqn_pixel_atari(name):
     # config.network_fn = lambda optimizer_fn: DuelingNatureConvNet(config.history_length, n_actions, optimizer_fn)
     config.policy_fn = lambda: GreedyPolicy(epsilon=1.0, final_step=1000000, min_epsilon=0.1)
     config.replay_fn = lambda: Replay(memory_size=1000000, batch_size=32, dtype=np.uint8)
+    config.reward_shift_fn = lambda r: np.sign(r)
     config.discount = 0.99
     config.target_network_update_freq = 10000
     config.max_episode_length = 0
@@ -104,6 +105,7 @@ def async_pixel_atari(name):
     # config.worker = OneStepSarsa
     # config.worker = NStepQLearning
     config.worker = OneStepQLearning
+    config.reward_shift_fn = lambda r: np.sign(r)
     config.discount = 0.99
     config.target_network_update_freq = 10000
     config.max_episode_length = 10000
@@ -123,6 +125,7 @@ def a3c_pixel_atari(name):
     config.optimizer_fn = lambda params: torch.optim.Adam(params, lr=0.0001)
     config.network_fn = lambda: OpenAIActorCriticConvNet(
         config.history_length, task.env.action_space.n, LSTM=True)
+    config.reward_shift_fn = lambda r: np.sign(r)
     config.policy_fn = SamplePolicy
     config.worker = AdvantageActorCritic
     config.discount = 0.99
diff --git a/utils/config.py b/utils/config.py
index 132a866..e7fa540 100644
--- a/utils/config.py
+++ b/utils/config.py
@@ -37,6 +37,7 @@ class Config:
         self.noise_decay_interval = 0
         self.target_network_mix = 0.001
         self.action_shift_fn = lambda a: a
+        self.reward_shift_fn = lambda r: r
         self.reward_weight = 1
         self.hybrid_reward = False
         self.target_type = self.q_target