Voyager is an open-ended embodied agent powered by Large Language Models (LLMs) that continuously explores, acquires new abilities, and makes novel discoveries in the Minecraft environment without human intervention. The Voyager system is primarily made up of three components. First, an 'automatic curriculum' guides the system's exploration. This curriculum is determined based on the system's progress and state, with an overarching goal of discovering a diverse array of objects and features. Second, a 'skill library' stores and retrieves complex behaviors. Each acquired skill is indexed by the embedding of its description, which is later used to retrieve that skill when faced with similar situations, and the development of such skills is also vital in minimizing catastrophic forgetting. Third, an 'iterative prompting mechanism' generates executable code for control using the environment's feedback, execution errors, and self verification. Voyager interacts primarily through blackbox queries with a Large Language Model (LLMs). For its action space, the system uses code rather than low-level motor commands since the former can easily represent temporally extended actions and compositional tasks, which are necessary for numerous long-term tasks in Minecraft. Generally, Voyager can establish unique tasks based on its current skill level and state of the world, improve skills based on environmental feedback, commit skills to memory for future similar tasks and explore the world in a self-sufficient way, continually seeking new tasks to complete.
