Hyperparameters used in SERL experiments (#55)

[OliEfr]

@Leo428 @youliangtan

Hi! Great work on the serl project and thanks for publishing it :) I'd also have a small question regarding the hyperparameters.

• I noticed that for instance for the examples/async_pcb_insert_drq task the max_steps is set to 1_000_000. Given that the robot env runs at 10Hz, this would mean that data collection takes at least 100_000s = 27hrs. Am I missing something here or do I need to lower the parameters to train in, say, 1 hr according to what is in the paper?
• Using the same 4090 GPU as in the paper for me the training batch_size=256 appears to be to big and I go OOM, but for 128 it works as intended!

Thank you so much!

Refers to #55

@lukahofer

[youliangtan]

• the 100000 is an arbitrary number we selected, in practice a significant lower buffer size is suffice
• yeap 128 batchsize is duffice too.