search

rainorangelemon / gnn-motion-planning

The official repo for NeurIPS 2021 paper 'Reducing Collision Checking for Sampling-Based Motion Planning Using Graph Neural Networks'
MIT License
93 stars 18 forks source link

Guidance to re-train explorer and smoother model #3

Closed wenjinzhang closed 1 year ago

wenjinzhang commented 1 year ago

Hi

Thank you very much for sharing your work. Would you please provide guidance to re-train the explorer and smoother models? I am also interested in how to generate the training dataset. Thank you in advance.

Best regards, Wenjin

rainorangelemon commented 1 year ago

Hello Wenjin,

Thanks for asking!

Please take a look at train_explorer.py and train_smoother.py for the training procedure. The code currently is a bit messy, so I do want to refactor them in the future. Please stay tuned.

To generate the dataset, I first generate a bunch of cases where the obstacles and the configurations of the robot are randomly sampled. Then, I use the BIT planner with a much longer timeout (e.g., 10 mins), to find the test cases which contain feasible solutions. These cases where BIT with longer timeout succeeds are saved as the dataset (including the training and test dataset). I will try my best to find the codes where I generate these datasets, but I cannot guarantee since I put these codes separately somewhere on a computer that I haven't used for nearly 1.5 years... But I hope you can get a sense about how I generate the dataset based on the description above.

Let me know if there are any other questions! I'd love to help.

wenjinzhang commented 1 year ago

Hi,

Thank you very much for the explanation. After checking train_explorer.py and train_smoother.py, I got it. I also got how to generate the dataset. This makes sense to me. I understand it is hard to find the old code. Hopefully, you can find it and post it on your repo. We appreciate your work.

Best regards, Wenjin