diff --git a/serl_launcher/serl_launcher/data/serl_memory_efficient_replay_buffer.py b/serl_launcher/serl_launcher/data/serl_memory_efficient_replay_buffer.py
index 1b6e15bd..0ae6052c 100644
--- a/serl_launcher/serl_launcher/data/serl_memory_efficient_replay_buffer.py
+++ b/serl_launcher/serl_launcher/data/serl_memory_efficient_replay_buffer.py
@@ -24,12 +24,12 @@ def __init__(
         for pixel_key in self.pixel_keys:
             pixel_obs_space = observation_space.spaces[pixel_key]
             if self._num_stack is None:
-                self._num_stack = pixel_obs_space.shape[-1]
+                self._num_stack = pixel_obs_space.shape[0]
             else:
-                assert self._num_stack == pixel_obs_space.shape[-1]
-            self._unstacked_dim_size = pixel_obs_space.shape[-2]
-            low = pixel_obs_space.low[..., 0]
-            high = pixel_obs_space.high[..., 0]
+                assert self._num_stack == pixel_obs_space.shape[0]
+            self._unstacked_dim_size = pixel_obs_space.shape[-1]
+            low = pixel_obs_space.low[0]
+            high = pixel_obs_space.high[0]
             unstacked_pixel_obs_space = Box(
                 low=low, high=high, dtype=pixel_obs_space.dtype
             )
@@ -71,13 +71,13 @@ def insert(self, data_dict: DatasetDict):
         if self._first:
             for i in range(self._num_stack):
                 for pixel_key in self.pixel_keys:
-                    data_dict["observations"][pixel_key] = obs_pixels[pixel_key][..., i]
+                    data_dict["observations"][pixel_key] = obs_pixels[pixel_key][i]
 
                 self._is_correct_index[self._insert_index] = False
                 super().insert(data_dict)
 
         for pixel_key in self.pixel_keys:
-            data_dict["observations"][pixel_key] = next_obs_pixels[pixel_key][..., -1]
+            data_dict["observations"][pixel_key] = next_obs_pixels[pixel_key][-1]
 
         self._first = data_dict["dones"]
 
@@ -151,12 +151,13 @@ def sample(
                 obs_pixels, self._num_stack + 1, axis=0
             )
             obs_pixels = obs_pixels[indx - self._num_stack]
+            obs_pixels = obs_pixels.transpose((0, 4, 1, 2, 3))
 
             if pack_obs_and_next_obs:
                 batch["observations"][pixel_key] = obs_pixels
             else:
-                batch["observations"][pixel_key] = obs_pixels[..., :-1]
+                batch["observations"][pixel_key] = obs_pixels[:, :-1, ...]
                 if "next_observations" in keys:
-                    batch["next_observations"][pixel_key] = obs_pixels[..., 1:]
+                    batch["next_observations"][pixel_key] = obs_pixels[:, 1:, ...]
 
         return frozen_dict.freeze(batch)