diff --git a/README.md b/README.md
index 41daaf121..6bfac65e2 100644
--- a/README.md
+++ b/README.md
@@ -74,6 +74,7 @@ sudo apt install libboost-all-dev
 ```
 git clone https://github.com/HorizonRobotics/alf
 cd alf
+pip install pybind11
 pip install -e . --extra-index-url https://download.pytorch.org/whl/cu118
 ```
 
diff --git a/alf/utils/tensorrt_utils.py b/alf/utils/tensorrt_utils.py
index d2fb21e76..7c04d5533 100644
--- a/alf/utils/tensorrt_utils.py
+++ b/alf/utils/tensorrt_utils.py
@@ -41,12 +41,9 @@
 
 # How to install dependencies (in a virtual env) on the deployment machine:
 # ```bash
-#   pip install onnx
+#   pip install onnx>=1.16.2 protobuf==3.20.2
 #
-#   pip install tensorrt
-# After this, tensorrt .so files will be installed in your virtual env: .../site-packages/tensorrt_lib/
-# You need to copy all .so files under it to /usr/local/cuda/targets/x86_64-linux/lib/
-# Or alternatively, you can install it system-wide from https://developer.nvidia.com/tensorrt-getting-started
+#   pip install tensorrt>=10.0
 
 # For cuda 11.x,
 #   pip install onnxruntime-gpu
@@ -61,6 +58,11 @@
 # ['TensorrtExecutionProvider', 'CUDAExecutionProvider', 'CPUExecutionProvider']
 # The order of elements represents the default priority order of Execution Providers from highest to lowest.
 
+# NOTE: If onnxruntime complains about not finding libnvinfer.so or other tensorrt libs,
+# they can be found in your virtual env: .../site-packages/tensorrt_lib/
+# You need to copy all .so files under it to /usr/local/cuda/targets/x86_64-linux/lib/
+# Or alternatively, you can install it system-wide from https://developer.nvidia.com/tensorrt-getting-started
+
 
 def is_onnxruntime_available():
     return backend is not None
diff --git a/setup.py b/setup.py
index bc4724572..daabbe0ad 100644
--- a/setup.py
+++ b/setup.py
@@ -37,7 +37,7 @@
         'pathos==0.2.4',
         'pillow>=8',
         # 'procgen>=0.10.4',
-        'protobuf==3.20.1',
+        'protobuf==3.20.2',
         'psutil==5.9.8',
         'pybullet==2.5.0',
         'pyglet==1.3.2',  # higher version breaks classic control rendering