Are the positions of subgoals(in craft for example, iron, grass, wood, gold, ...) Fixed?

[heyzude]

May I ask Are the positions of subgoals fixed in all experiment environments, at multitask and zero-shot settings?
At first after reading the paper only, I thought that subpolicy pi is only given s_i (state), so the model can't treat variable subgoal position environments. I read the code and guess the positions are fixed in the code too.

Sorry for my question at this late time(2021).
Thank you in advance :)

[jacobandreas]

Hi---sorry about the very slow reply to this email! Objects move around between restarts; the "state" is not an indicator on (x, y) position, but instead an observation of a (randomly generated) environment that includes information about the agent's current location. See https://github.com/jacobandreas/psketch/blob/master/worlds/craft.py#L190.

…

On Fri, May 21, 2021 at 8:23 AM heyzude ***@***.***> wrote: May I ask Are the positions of subgoals fixed in all experiment environments, at multitask and zero-shot settings? At first after reading the paper only, I thought that subpolicy pi is only given s_i (state), so the model can't treat variable subgoal position environments. I guess the positions are fixed in the code too. Sorry for my question at this late time(2021). Thank you in advance :) — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#10>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AABZUTZS44VOM5BVGGFRGMTTOZGCVANCNFSM45JBSPTQ> .