Upon finishing bachelor's at Cornell University, I joined AgileSoDA as a summer intern. I was lucky enough to participate in a PoC to gain a real experience collaborate with an industry partner. The As-Is process was performed by an expert, but was a tedious task. The main goal was to design a RL algorithm for choosing pin ejector location for TV manfuacturing process. Using 3D CAD data, we designed the RL algorithm and utilized ParaView.
The main skills I used in this project are : PyTorch, Open-cv, 3D-CAD, NNI ParaView.
The data types that were given as in input were .vtu and .stl files. In order to read those data types, I had to research an open source tool and found ParaView as it provided the python API. Using that, I was able to preprocess the data and convert them into 2D images of the product with the help of Open-cv.
After preprocessing the data, I was responsible for RL training. RL training's quality highly depends on the neural network parameters. After researching open source yet again, I found and utilized NNI, Neural Network Intellgience. NNI is widely used for hyperparameter optimization.
After the training is finished, it is crucial to convert the result into a format that the end-user is able to use. The .stl file had to be the final output, so I worked on converting the RL data into .stl file for easy use. I also utilized ParaViewWeb Visualizer to view results in the local web browser.
Note that I cannot disclose further information due to privacy concerns.