“…RL-based agents are sometimes intrinsically motivated (Forestier et al, 2017;Colas et al, 2020;Akakzia et al, 2021;Hill et al, 2021). They imitate behaviors (Chevalier-Boisvert et al, 2019;Lynch and Sermanet, 2021), use hierarchical abstractions to decompose a complex task into simpler tasks (Oh et al, 2017;Eppe et al, 2019), and some of them can be trained with language to follow instructions (Hermann et al, 2017;Oh et al, 2017;Chaplot et al, 2018;Narasimhan et al, 2018;Chevalier-Boisvert et al, 2019;Hill et al, 2019Hill et al, , 2020Hill et al, , 2021Jiang et al, 2019;Colas et al, 2020).…”