The paper introduces SemTra, a framework for zero-shot adaptation of semantic skills in cross-domain settings. Semantic skills are interpretable experts’ behavior patterns. SemTra uses multi-modal models to extract skills from user inputs and a pre-trained language model to adapt these skills to the target domain. The framework uses a two-level hierarchy for adaptation: task and skill. Task adaptation transforms the extracted skills into a semantic skill sequence, while skill adaptation optimizes each semantic skill for the target domain context. The framework was evaluated using various environments, demonstrating its effectiveness in performing long-horizon tasks and adapting to different domains.
Publication date: 13 Feb 2024
Project Page: Not Provided
Paper: https://arxiv.org/pdf/2402.07418
