Is there any API to pull natural language description of each task and environment?

[MonumentalCloud]

Is there any API to pull natural language description of each task and environment?
and also,
how do you train in MT10? do you just train on single goal position for 10 environments? or should the 10 train environments also have different goal tasks?

Thank you in advance!

[csimo005]

As far as I know, no there is no api to get a natural language description of each task. Writing these description yourself would be fairly straight forward, then you could implement an env wrapper that makes this information accessible to the rest of your code.

I'm a bit confused by the second part of your question. I assume if you're targeting MT10 your trying to train a multi-task model, in which case you would train one model for all 10 tasks, no?

[MonumentalCloud]

Thanks for your thoughtful reply.

Yeah sorry about the confusion, I think I was confused myself asking this question, now I see that there are different goal position within the 10 environments in the MT10.

The only reason I asked for the API is that the natural language description in the original seems to be lacking in information quantity the agents might be able to take advantage of, and if everyone writes their own description, there might be standardization issue when it comes to writing papers.

[csimo005]

If there's enough a demand for it, maybe standardizing would be a good idea. I think I remember seeing some discussion about including a standard visual output would have to be done carefully, to not violate the meta learning setting. Similar care would probably need to be taken with a standardized language description api. You could consider opening a issue with a feature request to get more attention for the idea.