Suppose we use DRAGO to test a ten-metaworld-task sequence like "continual world benchmark" to evaluate its ability to resist forgetting. Is it necessary to load the checkpoint of the previous task and apply the reviewer at the beginning of each task as per your method? for example, is it necessary to load ckpt of task1 befor running task2 ？I would like to understand your original intention when you were designing it. Thanks！！
Suppose we use DRAGO to test a ten-metaworld-task sequence like "continual world benchmark" to evaluate its ability to resist forgetting. Is it necessary to load the checkpoint of the previous task and apply the reviewer at the beginning of each task as per your method? for example, is it necessary to load ckpt of task1 befor running task2 ？I would like to understand your original intention when you were designing it. Thanks！！