548a482d4496ce109cddfbeae5defa7d-Supplemental-Conference.pdf

In this section, we provide PIE-G's detailed settings. In this paper, we use the ready-mademodels from the following link: ResNet: https://github.com/pytorch/vision; PIE-G with MoCo-v2 also achieves competitive sample efficiency. The episode lengths in Drawer World tasks are 200 steps with 4 action repeat. Table 9: Compare with RRL.RRL barely generalize to the new environmentsintheDMC-GB.