JEDI: Joint-Embedding Demonstration Compression for In-Context Imitation Learning
JEDI: Joint-Embedding Demonstration Compression for In-Context Imitation Learning
This research proposes JEDI, a method that compresses demonstrations into latent slots trained to predict future keyframes in a frozen video-JEPA space. Task and process counterfactuals help preserve execution details, and experiments on LIBERO and real robots show improved success and demonstration following over action-only compression.
This paper introduces JEDI, a method for compressing demonstrations in imitation learning by encoding them into latent slots that predict future keyframes. The approach uses task and process counterfactuals to retain execution details and demonstrates improved performance on LIBERO and real robots compared to action-only compression.
Source: arXiv Robotics — research abstracts · Read original article ↗
Loading article text…
Source：arXiv Robotics — research abstracts · arxiv.org