Reproduce “JEDI: Joint-Embedding Demonstration Compression for In-Context Imitation 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
Jiazhao Liang, Zhewen He, Shuaihang Yuan, Hao Huang, Yi Fang
JEDI compresses each robot demonstration into eight cached latent slots trained to predict the next keyframe representation from a query robot’s current observation. Task and process negatives encourage slots to retain execution details. In LIBERO-Order, it follows demonstrated object order in 84.1% of rollouts, versus 52.2% for a matched action-only compressor; evidence is bounded to tested tasks and setup.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.