Reproduce “Why Koopman Structure Does Not Help Differentiable World Models for Humanoid Whole-Body Tracking” 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
Haizhou Ge, Mengshuo Wang, Yufei Jia, Lei Han, Lu Shi, Guyue Zhou, Ruqi Huang
The authors test whether a Koopman world model improves Unitree G1 tracking, replacing an MLP in a fixed training pipeline. With state re-lifting at each imagined step, the model constrains predicted increments to be affine in action. Returns are near parity (0.830 versus 0.823), but Koopman predictions are less accurate and iterations take about 15% longer in this simulated setup.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.