Voyager

Collection Summary

Minecraft에서 automatic curriculum, executable skill library, iterative prompting을 사용해 open-ended lifelong learning을 수행한다.

Evolution History Store

• **Uses rollout buffer**: no generic trajectory replay buffer.
• **Persistent equivalent**: ever-growing skill library of executable code indexed by description embeddings.
• **Reuse path**: relevant skills are retrieved and composed for new tasks; environment errors and self-verification refine programs before admission.
• **Security relevance**: malicious skill admission, code-supply-chain compromise, embedding retrieval collision, unsafe composition, evaluator bypass, executable persistence.