Reproduce “SaiVLA-JEPA: Evaluating Future Latent Prediction in JEPA-Conditioned VLA Control” 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
Rufeng Xiao, Ruihao Xiao, Wei Chen, Xinhai Sun
The authors test whether future latent prediction improves robot manipulation beyond current visual conditioning. FutureJEPA adds training-only prediction of frozen visual representations to a policy while keeping its action route unchanged. In paired LIBERO-Plus episodes with 300-step caps, it succeeds on 7,924 of 10,030, versus 7,921 for CurrentJEPA, a difference not statistically detectable in this single training run.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.