3 ms·
Upvoting this because I love innovation in this space, and I use both ROS and LLMs daily. Kudos on the work here, I can see value for developers trying to break
by nathanfig 3y ago
Upvoting this because I love innovation in this space, and I use both ROS and LLMs daily. Kudos on the work here, I can see value for developers trying to break into ROS in particular.
That being said, and I might just need to try it out to understand, but it's not clear to my why this tool would be better than just talking to ChatGPT directly. Particularly GPT-4. There's the guided workflow aspect, but even that I could probably emulate by copying the description from this article and asking GPT-4 to do the same. It might just require more manual copying and pasting.
- RoboticTech 3y agoThank you. One of main features of ROScribe (which is still under development) is the use of RAG (Retrieval Augmented Generation) for robot software generation. In RAG, the code, rather than being generated by the LLM, is pulled from ROS index by LLM (LLM keeps all documentations and meta data of ROS index in a vector data base to figure out what piece of code is most suitable for the task at hand.) We are working on this part and we plan to publish a conference paper on it. So the ultimate solution is a mixture of codes pulled from online repositories (ROS index), internal data (companies own IPs), and the code that LLM generates. LLM is handling all of three segments, but only codes as minimum as needed.
- nathanfig 3y agoThat is cool, and especially useful since LLMs often get details like URLs to repositories wrong, so I could see how this system would increase the accuracy of the output.