fix: OpenVINO optimization notebook deployment flow in 003 notebook

[zhuo-yoyowz]

Summary

Fixed issues in 003_OpenVINO_Optimization.ipynb, including:

1. Fixed git clone Physical AI commit number, since the repo is developed fast and the last version used API out of date, causing errors during running.
2. Removed the hard-coded robot calibration file path. Remind users to manually config the file path according to their env.
3. Added a reminder for user to give permission to USB cameras ports in Linux OS.
4. Added two choices for Model Loading, either using exported model already in local env or downloading an open-sourced pi-0.5 fine-tuned model.
5. Updated some APIs according to the updated version of Physical API master branch.

[maxxgx]

Thanks for the PR.

Do you find these tutorials helpful? I was thinking we could reduce them, as they have quite a bit of overlap. For example, the collect_train_deploy notebook covers most of the topics.

Also, I believe it will be hard to keep this branch up to date, since breaking changes are quite regular while we're still in early development.

[maxxgx]

I think these tutorial were written before release. Now there is no need to clone the physicalai repo since it's already available on pypi, so you could pin a released version

[zhuo-yoyowz]

Yes, just found the formal release version. Will modify the commit number to the released version.

[zhuo-yoyowz]

Thanks for the PR.

Do you find these tutorials helpful? I was thinking we could reduce them, as they have quite a bit of overlap. For example, the collect_train_deploy notebook covers most of the topics.

Also, I believe it will be hard to keep this branch up to date, since breaking changes are quite regular while we're still in early development.

Thank you for your comments!
We do think Jupyter notebooks are helpful to developers, especially when they try to get hands-on to a whole new set of tools.
I tested all notebooks. Of course, 001 and 002 are documentations actually, therefore, in the future, maybe we could move them to Readme, so that developers could read them directly when navigating to the "tutorials" path?
004 notebook provides good pure simulation sample codes, which I think is also critical, especially to developers who currently don't own any physical robotic arms to test, but want to see how's everything working, whether it's hard to run and some initial performance results.
003 shows good optimization and deployment codes, but need some fixes, which I think could be also helpful when developers download an open-sourced model and would like to connect the robotic arms to run some initial inferences.
005collect_train_deploy does contain a full e2e process, however, it needs at least an A100 to run from the beginning, i.e. the model training. If only keeps this one notebook, developers could run nothing but have to leave the webpage soon if they don't own an A100.
So I hope these notebooks could be merged to the master branch, and we could provide clear guidance on how to use these notebooks to get started quickly and easily. I'm happy to be part of this for maintaining and upgrading, also writing tech blogs to promote these.

[maxxgx]

Thanks for the PR.
Do you find these tutorials helpful? I was thinking we could reduce them, as they have quite a bit of overlap. For example, the collect_train_deploy notebook covers most of the topics.
Also, I believe it will be hard to keep this branch up to date, since breaking changes are quite regular while we're still in early development.

Thank you for your comments! We do think Jupyter notebooks are helpful to developers, especially when they try to get hands-on to a whole new set of tools. I tested all notebooks. Of course, 001 and 002 are documentations actually, therefore, in the future, maybe we could move them to Readme, so that developers could read them directly when navigating to the "tutorials" path? 004 notebook provides good pure simulation sample codes, which I think is also critical, especially to developers who currently don't own any physical robotic arms to test, but want to see how's everything working, whether it's hard to run and some initial performance results. 003 shows good optimization and deployment codes, but need some fixes, which I think could be also helpful when developers download an open-sourced model and would like to connect the robotic arms to run some initial inferences. 005collect_train_deploy does contain a full e2e process, however, it needs at least an A100 to run from the beginning, i.e. the model training. If only keeps this one notebook, developers could run nothing but have to leave the webpage soon if they don't own an A100. So I hope these notebooks could be merged to the master branch, and we could provide clear guidance on how to use these notebooks to get started quickly and easily. I'm happy to be part of this for maintaining and upgrading, also writing tech blogs to promote these.

Thank you for your feedback. I think we still improvements before merging it to main branch:

• 002 can be removed and we should point users to the PhysicalAI Studio documentation since it has released an overview of the application: https://github.com/open-edge-platform/physical-ai-studio/blob/main/application/README.md
• I find the code in 003 and 004 not "clean" and often deviates from the recommended usage shown in the repo. It looks like that several shortcuts are used to make the code work.
• The hugging face links points to personal repos, we can use the official collection instead: https://huggingface.co/collections/OpenVINO/physical-ai
• If would be amazing if we could use a proper simulation environment like mujoco running on the browser: https://lerobot.binhph.am/

[zhuo-yoyowz]

Thanks for the PR.
Do you find these tutorials helpful? I was thinking we could reduce them, as they have quite a bit of overlap. For example, the collect_train_deploy notebook covers most of the topics.
Also, I believe it will be hard to keep this branch up to date, since breaking changes are quite regular while we're still in early development.

Thank you for your comments! We do think Jupyter notebooks are helpful to developers, especially when they try to get hands-on to a whole new set of tools. I tested all notebooks. Of course, 001 and 002 are documentations actually, therefore, in the future, maybe we could move them to Readme, so that developers could read them directly when navigating to the "tutorials" path? 004 notebook provides good pure simulation sample codes, which I think is also critical, especially to developers who currently don't own any physical robotic arms to test, but want to see how's everything working, whether it's hard to run and some initial performance results. 003 shows good optimization and deployment codes, but need some fixes, which I think could be also helpful when developers download an open-sourced model and would like to connect the robotic arms to run some initial inferences. 005collect_train_deploy does contain a full e2e process, however, it needs at least an A100 to run from the beginning, i.e. the model training. If only keeps this one notebook, developers could run nothing but have to leave the webpage soon if they don't own an A100. So I hope these notebooks could be merged to the master branch, and we could provide clear guidance on how to use these notebooks to get started quickly and easily. I'm happy to be part of this for maintaining and upgrading, also writing tech blogs to promote these.

Thank you for your feedback. I think we still improvements before merging it to main branch:

• 002 can be removed and we should point users to the PhysicalAI Studio documentation since it has released an overview of the application: https://github.com/open-edge-platform/physical-ai-studio/blob/main/application/README.md
• I find the code in 003 and 004 not "clean" and often deviates from the recommended usage shown in the repo. It looks like that several shortcuts are used to make the code work.
• The hugging face links points to personal repos, we can use the official collection instead: https://huggingface.co/collections/OpenVINO/physical-ai
• If would be amazing if we could use a proper simulation environment like mujoco running on the browser: https://lerobot.binhph.am/

Thanks for the feedback.

I updated 003 to use the released PyPI package flow end-to-end instead of relying on a cloned local physicalai repo. The hardware extras are now installed from the pinned release as well.

I also replaced the personal Hugging Face repos with the official OpenVINO Physical AI collection model (OpenVINO/pi05-libero-fp16-ov) and added a guard to make clear that the official hub model is for model-loading validation unless the user verifies that its robot/camera/action schema matches their real SO101 setup.

For the broader 004 cleanup and browser-based simulation direction, I'll submit new PR to fix the issues. For the 002 notebook suggestion, yes, I think we could remove it and point users to documentation as you suggested. But for now, I think we need to show the tutorials repo first and could improve it step by step, so maybe this notebook could be removed later?