When I evaluate task 'open_fridge', I notice that even after calling the reset_to_demo function, the initial object pose configurations sometimes (although not always) remain different compared to the data. In this situation, I observe that the init_episode function attempts different numbers of times. I hypothesize that the sampling-based planner used in RLBench introduces a degree of randomization, which leads to different outcomes in the task validation process. Have you made similar observations? If so, do you have any suggestions on how this issue might be resolved?
Hi,
Thank you for your wonderful implementation.
When I evaluate task 'open_fridge', I notice that even after calling the reset_to_demo function, the initial object pose configurations sometimes (although not always) remain different compared to the data. In this situation, I observe that the init_episode function attempts different numbers of times. I hypothesize that the sampling-based planner used in RLBench introduces a degree of randomization, which leads to different outcomes in the task validation process. Have you made similar observations? If so, do you have any suggestions on how this issue might be resolved?