no docker on jetson

[legned]

Hi zz990099,

Thank you for your contribution. I want to ask if you deploy the model on Jetson directly without Docker? I am using ROS2 to manipulate robotics, so if a Docker in a ROS2 node is not appropriate, I think. What kind of environment should I install instead of Docker? Thank you

[zz990099]

Hi! Although I haven't personally deployed this directly on a Jetson host, the setup should work as long as you configure your host environment according to the specifications in the Dockerfile (link).

For foundationpose_cpp, the following dependencies are required:

​​1. glog​​, ​​gtest, eigen3, assimp, opencv ...​​

​​2. CMake 3.22+​​ (necessary for locating CUDAToolkit.cmake)

​​3. CV-CUDA​​

​​4. TensorRT 8.6+​​ (recommended: ​​TensorRT 10+​​)

​​5. Ubuntu 22.04​​ is strongly advised (Note: JetPack 6 is preferred since JetPack 5 is based on Ubuntu 20.04).

Good luck with your deployment!

[legned]

Hi! Although I haven't personally deployed this directly on a Jetson host, the setup should work as long as you configure your host environment according to the specifications in the Dockerfile (link).

For foundationpose_cpp, the following dependencies are required:

​​1. glog​​, ​​gtest, eigen3, assimp, opencv ...​​

​​2. CMake 3.22+​​ (necessary for locating CUDAToolkit.cmake)

​​3. CV-CUDA​​

​​4. TensorRT 8.6+​​ (recommended: ​​TensorRT 10+​​)

​​5. Ubuntu 22.04​​ is strongly advised (Note: JetPack 6 is preferred since JetPack 5 is based on Ubuntu 20.04).

Good luck with your deployment!

Thank you for your kind reply. I will update after finishing deployment if successful.

[legned]

Hi! Although I haven't personally deployed this directly on a Jetson host, the setup should work as long as you configure your host environment according to the specifications in the Dockerfile (link).

For foundationpose_cpp, the following dependencies are required:

​​1. glog​​, ​​gtest, eigen3, assimp, opencv ...​​

​​2. CMake 3.22+​​ (necessary for locating CUDAToolkit.cmake)

​​3. CV-CUDA​​

​​4. TensorRT 8.6+​​ (recommended: ​​TensorRT 10+​​)

​​5. Ubuntu 22.04​​ is strongly advised (Note: JetPack 6 is preferred since JetPack 5 is based on Ubuntu 20.04).

Good luck with your deployment!

Hi

I downloaded the ONNX model from https://nvidia-isaac-ros.github.io/repositories_and_packages/isaac_ros_pose_estimation/isaac_ros_foundationpose/index.html#quickstart with the command below. And found out that the input tensor names are ['input1', 'input2']�. What I downloaded from your Google Drive is ['render_input', 'transf_input']. Is there any difference between these two models?
�
wget 'https://api.ngc.nvidia.com/v2/models/nvidia/isaac/foundationpose/versions/1.0.0_onnx/files/refine_model.onnx' -O refine_model.onnx &&
wget 'https://api.ngc.nvidia.com/v2/models/nvidia/isaac/foundationpose/versions/1.0.0_onnx/files/score_model.onnx' -O score_model.onnx

[zz990099]

Hi! The model we placed on Google Drive is our own exported ONNX model, and its parameters may not be consistent with the model on NGC. Additionally, our model's support for different batch sizes is not as good as the model on NGC. If you modify the input and output names of the model in the code, it might be compatible with the model from NGC.

[legned]

Hi! The model we placed on Google Drive is our own exported ONNX model, and its parameters may not be consistent with the model on NGC. Additionally, our model's support for different batch sizes is not as good as the model on NGC. If you modify the input and output names of the model in the code, it might be compatible with the model from NGC.

Hi,

I have deployed without Docker on Jetson. But when I ran the program, the output was different from the original Python version output. Do you know if this is a bug?

[zz990099]

The param refine_itr at link is different from the python version which is 3 by defaut. You may try modify this and check the results.

[zz990099]

I have deployed without Docker on Jetson. But when I ran the program, the output was different from the original Python version output. Do you know if this is a bug?

Hi! This issue might be related to the oriented bounding box computation. We addressed this in PR #31. Feel free to try that fix if you're still encountering the problem.

[lqy668honor]

你好！虽然我还没有亲自在 Jetson 主机上直接部署过它，但只要你按照 Dockerfile（链接）中的规范配置主机环境，设置应该就可以正常工作。

对于foundationpose_cpp，需要以下依赖项：

​​1. glog​​、​​gtest、eigen3、assimp、opencv……​​

​​2. CMake 3.22+​​（用于定位 CUDAToolkit.cmake）

​​3. CV-CUDA​​

4. TensorRT 8.6+​​（推荐：​​TensorRT 10+ ​​）

5. 强烈建议使用 Ubuntu 22.04​​（注意：由于 JetPack 5 基于 Ubuntu 20.04，因此最好使用JetPack 6 ）。

祝您部署顺利！

您好！我已经在jetpack 6 系统的设备上无docker实现了这个项目。但最近在尝试把他迁移到jetpack 5.1系统的设备上，在onnx模型转engin模型的过程中出现了一些问题，想请问下您之前尝试过这种操作吗？具备可行性吗？

[zz990099]

@lqy668honor 您好！onnx转换engine的过程应该取决于TensorRT的版本，我尝试的过程中TensorRT8.6应该是可以的，8.5及以下会出现算子不支持的问题。

[lqy668honor]

@lqy668honor 您好！onnx转换engine的过程应该取决于TensorRT的版本，我尝试的过程中TensorRT8.6应该是可以的，8.5及以下会出现算子不支持的问题。

[10/27/2025-15:50:08] [W] [TRT] Skipping tactic 0x0000000000000000 due to exception Incompatible effective shapes in operation /pos_embed/Add(b_pw:add) between: /Transpose_2_output_0'-(f32[252,400,512][]so[]md[2,-1,-1]) and /pos_embed/Slice_output_0'-(f32[1,252,512][]so[]).
[10/27/2025-15:50:08] [W] [TRT] Skipping tactic 0x0000000000000000 due to exception Incompatible effective shapes in operation /pos_embed/Add(b_pw:add) between: /Transpose_2_output_0'-(f16[252,400,512][]so[]md[2,-1,-1]) and /pos_embed/Slice_output_0'-(f16[1,252,512][]so[]).
[10/27/2025-15:50:08] [W] [TRT] Skipping tactic 0x0000000000000000 due to exception Incompatible effective shapes in operation /pos_embed/Add(b_pw:add) between: /Transpose_2_output_0'-(f16[252,400,512][]so[]md[2,-1,-1]) and /pos_embed/Slice_output_0'-(f16[1,252,512][]so[]).
[10/27/2025-15:50:08] [E] Error[10]: [optimizer.cpp::computeCosts::3728] Error Code 10: Internal Error (Could not find any implementation for node {ForeignNode[/att/Constant_2_output_0.../ReduceMean]}.)
[10/27/2025-15:50:08] [E] Error[2]: [builder.cpp::buildSerializedNetwork::751] Error Code 2: Internal Error (Assertion engine != nullptr failed. )
[10/27/2025-15:50:08] [E] Engine could not be created from network
[10/27/2025-15:50:08] [E] Building engine failed
[10/27/2025-15:50:08] [E] Failed to create engine from model or file.
[10/27/2025-15:50:08] [E] Engine set up failed
没错，我的是tensorrt8.5.2,出现的就是上面的错误。好像是因为tensorrt 8.6以上版本才支持opset15,您给的onnx模型opset是17，而opset14到15之间的gap很大。我现在的想法是：我重新导出onnx模型，限制opset的版本为14会不会解决这个问题？请问您尝试过吗？