clvrai/sprint

Code and website for for SPRINT: Scalable Policy Pre-Training via Language Instruction Relabeling

sprint's Issues

Question about computing the reward in PickupObjectAction()

Hi,
"Hi, I'm evaluating the EVAL Instruction environment and noticed that in some tasks, the reward does not equal 1 even after the agent completes a skill.

For example, in "pick_two_obj_and_place-PotatoSliced-None-SinkBasin-13/trial_T20190908_181751_435129" with 9 skills required (in this line), the agent initially attempts to pick up a knife from the counter's corner.

There are two knives at this location (Figure below), and ideally, picking up either shoud get a reward of 1. However, the reward function only returns 1 when picking the correct one based on the expert's demonstration, i.e., the innermost knife (see this line), while the outermost knife is also possible (and in your code, the agent selected the outermost one since it's closer distance, in here).

So, I think the agent will always fail in this scene. Do you have any idea to fix this situation?

When is the implementation public?

Hi team:

Thank you for sharing the great work! When is the implementation be public?

Hyperparameter for pretraining primitive skills

Thank you for sharing the code.

I'm trying to pretrain primitive skills using IQL. I confused about the batch size you used for pretraining. In the paper, you mentioned bs=1024, however, in both code and released checkpoints, it's 64. So what exact batch size do you use for pretraining?

And, could you share the hardware spec you used for pretraining and how long it takes for pretraining?

Thank you.