Clarification on trainer and agent

[realjoshqsun]

Hello,

I am working in ML, but I am new to Habitat. I'm slightly confused about the purpose of some files: hierarchical_trainer.py, nonlearning_agents.py, and robo_vln_trainer.py.

From my understanding, these files import functions from Habitat, which I expect are used for training purposes, given their filenames include "_trainer". And "agent" is for evaluation. However, the README mentions an offline data buffer, suggesting that the trainers extract data from this buffer, without live interaction with the Habitat environment. This has led to some confusion.

Could you please clarify the following:

The specific roles and functionalities of hierarchical_trainer.py, nonlearning_agents.py, and robo_vln_trainer.py.
Whether my understanding that the trainers operate by extracting data from an offline buffer, thus not requiring Habitat for real-time simulation, is correct.
I might be missing some context or misunderstanding the architecture, so the suggestions you give would help me understand the code.

Thank you very much for your time and assistance.

Josh