Reproduce “Few-Shot Residual Adaptation of Pretrained Latent World Models” 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
Jakub Suliga, Pankhuri Vanjani, Nisarga Nilavadi, Wolfram Burgard, Rudolf Lioutikov
The authors adapt a visual model for box pushing under center-of-mass and friction shifts, keeping the controller fixed. They learn corrections from prediction errors inside the predictor and on predicted states over three episodes, without rewards or parameter estimates. On the real robot, success rose from 0.0% to 90.0% and from 3.3% to 100.0% under the tested shifts.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.