Why the eval success rate is decreasing?

[bbbbbbbbly]

This image show I ran command python train_dsrl.py --config-path=cfg/robomimic --config-name=dsrl_can.yaml after train 2 hours.

[ZhHe11]

I have successfully reproduced the results reported in the paper after training for 100 hours on our server. However, I have some concerns about the interaction efficiency. Is that expected?

Update:

Setting n_envs=50 for the can task works much better (similar to the DPPO settings).