Detail of maze2d experiment

[Super1ce]

I reproduced the results of the original paper on locomotion, but encountered some difficulties on maze2d. I would like to ask:

1. Is the score given in the original paper normalized?
2. The diffuser step given in the appendix only mentions locomotion and block-stacking, is maze2d also 100?
3. The paper says to use start and goal locations, is the goal location obtained through the environment’s get_target and reset_to_location? Is the start location referring to the current observation of the agent? Is the goal location placed at horizon - 1? After setting it up this way, the result I reproduced only got a non-normalized score of 87.

[jannerm]

Are you on the maze2d branch?

1. Yes
2. The number of diffusion steps are in the config file here
3. The goal location is randomized here

[Super1ce]

Sorry, I didn’t notice this branch. I will try to reproduce it again. Thank you.

[Super1ce]

Are you on the maze2d branch?

1. Yes
2. The number of diffusion steps are in the config file here
3. The goal location is randomized here

I saw a comment saying that the replan method can also be used during evaluation, but when I used the replan method of locomotion to replace the open-loop here, the return was all 0. Is it necessary to make additional changes to take effect?

[jannerm]

Closed-loop planning should work here; it will just be slower. What do the plans and rollouts look like?