Reproduce “Learning New Tasks via Reusable Skills: Skill-Compositional Experts for Embodied Continual Learning” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Shuaike Zhang, Shaokun Wang, Haoyu Tang, Jianlong Wu, Liqiang Nie
SCE addresses continual robot manipulation by segmenting demonstrations into reusable skills and adding separate experts for skill execution and transitions within a pretrained action decoder. On LIBERO-Long, it reports AUC 70.2, Final SR 73.9, and NBT 2.5, ahead of Task-MoE on these measures. It does not lead on forward transfer; these results do not establish a single cause.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.