Issue with goal condition example experiments

[www-Ye]

Hi, I'm wondering if the goal condition experiment examples are working correctly. I noticed that these experiment examples seem to be located in the 'temporarily_deactivated' folder.

[jakegrigsby]

Hi @www-Ye, yes the goal-conditioned examples have not been supported for a few weeks now. The hindsight relabeling technique was built on a convoluted sequence of gym.Wrappers that I removed as part of a large update (PR #51). I need to remake them with a new system. I've gone ahead and done this for the maze problem in the paper in a new example 13_mazerunner_relabeling (PR #74). I added a lot of comments so that it might serve as a copy/paste starting point for another env (or simpler HER).

Unfortunately I currently have no plans to remake the goal-conditioned crafter experiments, as the training runs take too long to test/debug. However I put a lot of effort into replicating those experiments in the git history. You can rewind to https://github.com/UT-Austin-RPL/amago/tree/v2 to use it if necessary, and there used to be a jupyter notebook demo where you could load a pre-trained checkpoint and visualize crafter instructions (https://github.com/UT-Austin-RPL/amago/blob/v2/examples/crafter_pixels_demo.ipynb).

[www-Ye]

Thank you for your response and the updates. I also noticed that I was able to successfully train using the goal condition mazerunner with the amago v2 version.
I was wondering why you chose to remove all of the old goal condition code. From the mazerunner experiments, what are the key differences between the two versions?