Reproduce “AD-WM: Action-Discriminative World Models for Counterfactual Model Predictive 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
Jiabin Qiu, Zixuan Chen, Hongye Cao, Jieqi Shi, Jing Huo, Yang Gao
The paper introduces AD-WM for robot control, training a world model to preserve action-dependent differences in predicted futures for MPC planning. It predicts residual changes and trains predicted transitions to recover actions using inverse dynamics and normalized recovery objectives; recovery heads are discarded at deployment. On OGBench-Cube hard-start evaluation, AD-WM reached 52.0 ± 3.1% success versus 3.7 ± 1.4% for matched LeWM, but residual prediction also contributed.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.