OpenBot

[Don Zagorski]

The open source "OpenBot" uses your smartphone + $50 worth of hardware on a 3D-printed frame.

High-level logic runs on Android, low-level motor interfaces are on Arduino.

https://www.youtube.com/watch?v=qc8hFLyWDOM&feature=youtu.be

Who wants to build one?

[Austin Keller]

Appendix C talks about the process for training it to navigate. They ended up only collecting 30 minutes of data by having someone steer the bot with a game controller through a few obstacle courses. I wonder if we could train a network for the pop can challenge this way (or at least part of it). A 1 million parameter model seems doable with Google Colab's free GPU resources too.

[Austin Keller]

I modified a notebook to test how long it'd take to train a similar sized model as long as they did (~1 million parameters). Using 100 epochs and a batch size of 128 (the upper end of what they tested) it trains in less than 15 minutes with the GPU runtime.

They have a nice README in the OpenBot repo that describes how to collect the data and even how to run the training job and recompile the app with the trained model. It seems very doable!

[Bill Adcock]

Austin
Which gpus can be used?  Nvidia?, amd?  ...

Bill

---

--
You received this message because you are subscribed to the Google Groups "SnoCo Robotics Club" group.
To unsubscribe from this group and stop receiving emails from it, send an email to snoco-robotic...@googlegroups.com.
To view this discussion on the web visit

https://groups.google.com/d/msgid/snoco-robotics/baf62444-0dd7-47a1-a00f-57de114e0ce7n%40googlegroups.com

.

[Austin Keller]

Tensorflow seems to only have built-in support for Nvidia. AMD is building out support for their GPUs with the tensorflow-rocm project, though. I only have experience with the cloud GPUs provided by Google colab. I assume those are Nvidia?

Austin