search

AIRLab-POLIMI / BTGenBot

BTGenBot: a system to generate behavior trees for robots using lightweight (~7 billion parameters) large language models (LLMs)
MIT License
52 stars 6 forks source link

I have a few questions. #4

Closed sea-hot closed 5 months ago

sea-hot commented 6 months ago

"Your assignment seems really interesting and impressive. So, I have a few questions.

  1. method 1) For the example_task, when a simple text command is input, 2) it produces results according to a predefined output format, 3) and for the assignment, if an explanation for the example_task, such as place, object, coordinates, etc., is provided, 4) is it correct that in fine-tuning, some of the output keywords in the output are changed?"

  2. I might have missed it, but is it correct that it is still impossible to design a behavior tree accordingly if an abstract command is input without precise information (no place) for the high-level task, and then supplemental explanations (place, object, coordinates) are inputted for the assignment?

  3. Or, is there a specific XML syntax that must be used whenever an abstract command is input?

  4. Sometimes, meaningless words follow, similar to photos. Does this mean it's not working properly? 20240401_233748

Your generosity has truly made a significant impact, and I am immensely thankful for everything.

Hi-Zed commented 6 months ago

You can find all the details of our approach in our paper here: https://arxiv.org/abs/2403.12761 Regarding the output in the image in point 4. Yes, something is not working properly because we never encountered a completely gibberish output. Sometimes the LLM gets stuck in an endless generation, but usually, it keeps the behavior tree syntax. I suggest you look at our example prompts here: https://github.com/AIRLab-POLIMI/BTGenBot/tree/master/prompt