Reproduce “PREM: Prediction-Regulated Memory for Robot Manipulation” 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

Xiao Huang, Ya Zhang, Weidi Xie

PREM adds a fixed-size recurrent visual state to a vision-language-action policy, using action-conditioned future prediction to train the state and regulate its updates. It compares arriving visual features with a cached prediction. On RMBench M(1), it achieves 83.8% macro-average success across five memory-dependent task families, versus 14.4% for π0.5; this full-framework comparison does not isolate prediction’s contribution.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.