diff --git a/robot_sf/sensor/sensor_fusion.py b/robot_sf/sensor/sensor_fusion.py
index de79078..129a1e2 100644
--- a/robot_sf/sensor/sensor_fusion.py
+++ b/robot_sf/sensor/sensor_fusion.py
@@ -7,6 +7,7 @@
 
 import numpy as np
 from gymnasium import spaces
+from collections import deque
 
 PolarVec2D = Tuple[float, float]
 
@@ -111,6 +112,10 @@ class SensorFusion:
     def __post_init__(self):
         # Initialize the number of steps to cache based on the LiDAR observation space
         self.cache_steps = self.unnormed_obs_space[OBS_RAYS].shape[0]
+        self.stacked_drive_state = np.zeros((self.cache_steps, 5), dtype=np.float32)
+        self.stacked_lidar_state = np.zeros((self.cache_steps, len(self.lidar_sensor())), dtype=np.float32)
+        self.drive_state_cache = deque(maxlen=self.cache_steps)
+        self.lidar_state_cache = deque(maxlen=self.cache_steps)
 
     def next_obs(self) -> Dict[str, np.ndarray]:
         """
@@ -135,7 +140,13 @@ def next_obs(self) -> Dict[str, np.ndarray]:
         next_target_angle = next_target_angle if self.use_next_goal else 0.0
 
         # Combine the robot speed and target sensor data into the drive state
-        drive_state = np.array([speed_x, speed_rot, target_distance, target_angle, next_target_angle])
+        drive_state = np.array([
+            speed_x,
+            speed_rot,
+            target_distance,
+            target_angle,
+            next_target_angle
+        ])
 
         # info: populate cache with same states -> no movement
         # If the caches are empty, fill them with the current states
@@ -147,19 +158,17 @@ def next_obs(self) -> Dict[str, np.ndarray]:
         # Add the current states to the caches and remove the oldest states
         self.drive_state_cache.append(drive_state)
         self.lidar_state_cache.append(lidar_state)
-        self.drive_state_cache.pop(0)
-        self.lidar_state_cache.pop(0)
-
-        # Stack the cached states into arrays
-        stacked_drive_state = np.array(self.drive_state_cache, dtype=np.float32)
-        stacked_lidar_state = np.array(self.lidar_state_cache, dtype=np.float32)
+        self.stacked_drive_state = np.roll(self.stacked_drive_state, -1, axis=0)
+        self.stacked_drive_state[-1] = drive_state
+        self.stacked_lidar_state = np.roll(self.stacked_lidar_state, -1, axis=0)
+        self.stacked_lidar_state[-1] = lidar_state
 
         # Normalize the stacked states by the maximum values in the observation space
         max_drive = self.unnormed_obs_space[OBS_DRIVE_STATE].high
         max_lidar = self.unnormed_obs_space[OBS_RAYS].high
         return {
-            OBS_DRIVE_STATE: stacked_drive_state / max_drive,
-            OBS_RAYS: stacked_lidar_state / max_lidar
+            OBS_DRIVE_STATE: self.stacked_drive_state / max_drive,
+            OBS_RAYS: self.stacked_lidar_state / max_lidar
             }
 
     def reset_cache(self):