Reproduce “La-JEPA: Compact Robot Policies with Latent World Modeling” 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
Mateo Clémente, George Z. Xue, Amin Abyaneh, Liheng Ma, Rui Heng Yang, Tongtong Cao, Zhanguang Zhang, Yingxue Zhang
La-JEPA trains compact robot-control transformers from scratch using pretrained V-JEPA 2.1 visual features, predicting future features alongside actions during training but generating only actions at inference. On corrected-instruction LIBERO-Plus, La-JEPA L reaches 88.0% success versus NowWAM’s 86.2%, with 1.78B versus 7.1B deployed parameters. La-JEPA B averages 60.8% on clean-trained RoboTwin, but remains below strongest embodied-pretrained methods on randomized scenes.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.