diff --git a/ding/policy/impala.py b/ding/policy/impala.py
index 4d693144bb7ed6637ca5a88aaf94b9be8801a39c..a37a1b257353025f8c1db03b03f81938a30dfb0f 100644
--- a/ding/policy/impala.py
+++ b/ding/policy/impala.py
@@ -39,7 +39,6 @@ class IMPALAPolicy(Policy):
                                                         | valid in serial training               | means more off-policy
         == ==================== ======== ============== ======================================== =======================
     """
-    unroll_len = 32
     config = dict(
         type='impala',
         cuda=False,
@@ -49,6 +48,8 @@ class IMPALAPolicy(Policy):
         priority=False,
         # (bool) Whether use Importance Sampling Weight to correct biased update. If True, priority must be True.
         priority_IS_weight=False,
+        # (int) the trajectory length to calculate v-trace target
+        unroll_len=32,
         learn=dict(
             # (bool) Whether to use multi gpu
             multi_gpu=False,
@@ -66,8 +67,6 @@ class IMPALAPolicy(Policy):
             discount_factor=0.9,
             # (float) additional discounting parameter
             lambda_=0.95,
-            # (int) the trajectory length to calculate v-trace target
-            unroll_len=unroll_len,
             # (float) clip ratio of importance weights
             rho_clip_ratio=1.0,
             # (float) clip ratio of importance weights
@@ -78,8 +77,6 @@ class IMPALAPolicy(Policy):
         collect=dict(
             # (int) collect n_sample data, train model n_iteration times
             n_sample=16,
-            # (int) the trajectory length to calculate v-trace target
-            unroll_len=unroll_len,
             # (float) discount factor for future reward, defaults int [0, 1]
             discount_factor=0.9,
             gae_lambda=0.95,
@@ -116,7 +113,7 @@ class IMPALAPolicy(Policy):
         self._learn_model = model_wrap(self._model, wrapper_name='base')
 
         self._action_shape = self._cfg.model.action_shape
-        self._unroll_len = self._cfg.learn.unroll_len
+        self._unroll_len = self._cfg.unroll_len
 
         # Algorithm config
         self._priority = self._cfg.priority
@@ -290,7 +287,6 @@ class IMPALAPolicy(Policy):
             Collect mode init method. Called by ``self.__init__``, initialize algorithm arguments and collect_model.
             Use multinomial_sample to choose action.
         """
-        self._collect_unroll_len = self._cfg.collect.unroll_len
         self._collect_model = model_wrap(self._model, wrapper_name='multinomial_sample')
         self._collect_model.reset()
 
diff --git a/dizoo/atari/config/serial/enduro/enduro_impala_config.py b/dizoo/atari/config/serial/enduro/enduro_impala_config.py
index 87f54bbc157464fb5b7365f2a0fe5ef4a89b7fd7..3e2874e8c4aa00dd817a37c2f1ac6d7a48b57e7c 100644
--- a/dizoo/atari/config/serial/enduro/enduro_impala_config.py
+++ b/dizoo/atari/config/serial/enduro/enduro_impala_config.py
@@ -14,6 +14,8 @@ enduro_impala_config = dict(
     ),
     policy=dict(
         cuda=True,
+        # (int) the trajectory length to calculate v-trace target
+        unroll_len=64,
         model=dict(
             obs_shape=[4, 84, 84],
             action_shape=9,
@@ -40,8 +42,6 @@ enduro_impala_config = dict(
             discount_factor=0.99,
             # (float) additional discounting parameter
             lambda_=1.0,
-            # (int) the trajectory length to calculate v-trace target
-            unroll_len=64,
             # (float) clip ratio of importance weights
             rho_clip_ratio=1.0,
             # (float) clip ratio of importance weights
@@ -52,8 +52,6 @@ enduro_impala_config = dict(
         collect=dict(
             # (int) collect n_sample data, train model n_iteration times
             n_sample=16,
-            # (int) the trajectory length to calculate v-trace target
-            unroll_len=64,
             # (float) discount factor for future reward, defaults int [0, 1]
             discount_factor=0.99,
             gae_lambda=0.95,
diff --git a/dizoo/atari/config/serial/pong/pong_impala_config.py b/dizoo/atari/config/serial/pong/pong_impala_config.py
index 7b5a536b7e978f958ef914b4903b6c2ce1fbdb9e..169b53d3cfc13fee43077c2898b1a6d1901f8839 100644
--- a/dizoo/atari/config/serial/pong/pong_impala_config.py
+++ b/dizoo/atari/config/serial/pong/pong_impala_config.py
@@ -15,6 +15,8 @@ pong_impala_config = dict(
     policy=dict(
         cuda=True,
         priority=False,
+        # (int) the trajectory length to calculate v-trace target
+        unroll_len=64,
         model=dict(
             obs_shape=[4, 84, 84],
             action_shape=6,
@@ -41,8 +43,6 @@ pong_impala_config = dict(
             discount_factor=0.9,
             # (float) additional discounting parameter
             lambda_=0.95,
-            # (int) the trajectory length to calculate v-trace target
-            unroll_len=64,
             # (float) clip ratio of importance weights
             rho_clip_ratio=1.0,
             # (float) clip ratio of importance weights
@@ -53,8 +53,6 @@ pong_impala_config = dict(
         collect=dict(
             # (int) collect n_sample data, train model n_iteration times
             n_sample=16,
-            # (int) the trajectory length to calculate v-trace target
-            unroll_len=64,
             # (float) discount factor for future reward, defaults int [0, 1]
             discount_factor=0.9,
             gae_lambda=0.95,
diff --git a/dizoo/atari/config/serial/qbert/qbert_impala_config.py b/dizoo/atari/config/serial/qbert/qbert_impala_config.py
index c957b7c3e6309d83956d0715f980ed561b2ed471..b206f43744be9ed41a9396a938d8659a9e1bc527 100644
--- a/dizoo/atari/config/serial/qbert/qbert_impala_config.py
+++ b/dizoo/atari/config/serial/qbert/qbert_impala_config.py
@@ -14,6 +14,8 @@ qbert_impala_config = dict(
     ),
     policy=dict(
         cuda=True,
+        # (int) the trajectory length to calculate v-trace target
+        unroll_len=64,
         model=dict(
             obs_shape=[4, 84, 84],
             action_shape=6,
@@ -40,8 +42,6 @@ qbert_impala_config = dict(
             discount_factor=0.9,
             # (float) additional discounting parameter
             lambda_=0.95,
-            # (int) the trajectory length to calculate v-trace target
-            unroll_len=64,
             # (float) clip ratio of importance weights
             rho_clip_ratio=1.0,
             # (float) clip ratio of importance weights
@@ -52,8 +52,6 @@ qbert_impala_config = dict(
         collect=dict(
             # (int) collect n_sample data, train model n_iteration times
             n_sample=16,
-            # (int) the trajectory length to calculate v-trace target
-            unroll_len=64,
             # (float) discount factor for future reward, defaults int [0, 1]
             discount_factor=0.9,
             gae_lambda=0.95,
diff --git a/dizoo/atari/config/serial/spaceinvaders/spaceinvaders_impala_config.py b/dizoo/atari/config/serial/spaceinvaders/spaceinvaders_impala_config.py
index 42d4434835d6e889471097eb26f4e169ec8e3e0e..dfa9c907d65db3f95b060352f24889f34ea63fb7 100644
--- a/dizoo/atari/config/serial/spaceinvaders/spaceinvaders_impala_config.py
+++ b/dizoo/atari/config/serial/spaceinvaders/spaceinvaders_impala_config.py
@@ -14,6 +14,8 @@ space_invaders_impala_config = dict(
     ),
     policy=dict(
         cuda=True,
+        # (int) the trajectory length to calculate v-trace target
+        unroll_len=64,
         model=dict(
             obs_shape=[4, 84, 84],
             action_shape=6,
@@ -40,8 +42,6 @@ space_invaders_impala_config = dict(
             discount_factor=0.9,
             # (float) additional discounting parameter
             lambda_=0.95,
-            # (int) the trajectory length to calculate v-trace target
-            unroll_len=64,
             # (float) clip ratio of importance weights
             rho_clip_ratio=1.0,
             # (float) clip ratio of importance weights
@@ -52,8 +52,6 @@ space_invaders_impala_config = dict(
         collect=dict(
             # (int) collect n_sample data, train model n_iteration times
             n_sample=16,
-            # (int) the trajectory length to calculate v-trace target
-            unroll_len=64,
             # (float) discount factor for future reward, defaults int [0, 1]
             discount_factor=0.9,
             gae_lambda=0.95,