diff --git a/metadrive/envs/base_env.py b/metadrive/envs/base_env.py
index 05174ec17..590190de3 100644
--- a/metadrive/envs/base_env.py
+++ b/metadrive/envs/base_env.py
@@ -566,6 +566,8 @@ def reset_sensors(self):
                 for name, sensor in self.engine.sensors.items():
                     if hasattr(sensor, "track") and name != "main_camera":
                         sensor.track(current_track_agent.origin, [0., 0.8, 1.5], [0, 0.59681, 0])
+        # Step the env to avoid the black screen in the first frame.
+        self.engine.taskMgr.step()
 
     def _get_reset_return(self, reset_info):
         # TODO: figure out how to get the information of the before step
diff --git a/metadrive/tests/test_functionality/test_first_frame_depth_cam.py b/metadrive/tests/test_functionality/test_first_frame_depth_cam.py
new file mode 100644
index 000000000..259a3790a
--- /dev/null
+++ b/metadrive/tests/test_functionality/test_first_frame_depth_cam.py
@@ -0,0 +1,70 @@
+import tqdm
+
+from metadrive.component.sensors.depth_camera import DepthCamera
+from metadrive.engine.asset_loader import AssetLoader
+from metadrive.envs.scenario_env import ScenarioEnv
+from metadrive.policy.replay_policy import ReplayEgoCarPolicy
+
+
+def test_first_frame_depth_cam():
+    env = ScenarioEnv(
+        {
+            # To enable onscreen rendering, set this config to True.
+            "use_render": False,
+
+            # !!!!! To enable offscreen rendering, set this config to True !!!!!
+            "image_observation": True,
+            "render_pipeline": False,
+
+            # ===== The scenario and MetaDrive config =====
+            "agent_policy": ReplayEgoCarPolicy,
+            "no_traffic": False,
+            "sequential_seed": True,
+            "reactive_traffic": False,
+            "start_scenario_index": 0,
+            "num_scenarios": 10,
+            "horizon": 1000,
+            "no_static_vehicles": False,
+            "vehicle_config": dict(
+                show_navi_mark=False,
+                use_special_color=False,
+                image_source="depth_camera",
+                lidar=dict(num_lasers=120, distance=50),
+                lane_line_detector=dict(num_lasers=0, distance=50),
+                side_detector=dict(num_lasers=12, distance=50)
+            ),
+            "data_directory": AssetLoader.file_path("nuscenes", unix_style=False),
+
+            # ===== Set some sensor and visualization configs =====
+            "daytime": "08:10",
+            "window_size": (800, 450),
+            "camera_dist": 0.8,
+            "camera_height": 1.5,
+            "camera_pitch": None,
+            "camera_fov": 60,
+            "interface_panel": ["semantic_camera"],
+            "sensors": dict(
+                # semantic_camera=(SemanticCamera, 1600, 900),
+                depth_camera=(DepthCamera, 800, 600),
+                # rgb_camera=(RGBCamera, 800, 600),
+            ),
+
+            # ===== Remove useless items in the images =====
+            "show_logo": False,
+            "show_fps": False,
+            "show_interface": True,
+        }
+    )
+
+    for ep in tqdm.trange(5):
+        env.reset()
+        for t in range(10000):
+            img = env.engine.get_sensor("depth_camera").get_image(env.agent)
+            assert not (img == 255).all()
+            if t == 5:
+                break
+            env.step([1, 0.88])
+
+
+if __name__ == '__main__':
+    test_first_frame_depth_cam()