Sub Task Identification [Deadline: 16.09]

[lauritowal]

Our agent must be able to succeed at a diverse set of subtasks
Each of these tasks could be viewed as a different goal which
requires a different policy to achieve the respective goal.
I assume that at evaluation time, the agent does not have knowledge about
which task environment it currently is in.

The agent will thus have to recognize the sub-tasks and then select the appropriate policy / actions.
It will use a classifier module (NN-based) for this meta-task, which we train on
(a subset of) the expert videos that we have for each task (environment) (e.g. via MineDojo Dataset)