About running the construction task.

[MurakamiNatsuki]

Hello!
Thank you for your interesting research!
I would like to try running the construction task, but could you please explain the method in a bit more detail?

[cnsdqd-dyb]

Thank you for your question. Recently, I've been attempting to fine-tune using reinforcement learning paradigms and planning to potentially open-source the fine-tuned small models, which caused me to overlook some issues.

Regarding the construction task, we provide 100 building blueprints that are described using blocks. Multiple agents analyze these blueprints and task instructions as input, then break down and execute the construction tasks. Their building process closely mirrors real construction processes, with strict constraints on block placement direction and position.

When running the construction task, we create multiple agents that join the server, along with a build_judger that enters in spectator mode to track and calculate task metrics. The task ends when either the metrics are completed or a timeout occurs.

This is a general overview of our task testing approach. If you have any specific questions about the details, please feel free to ask. Thank you!