Reproduce “MAPO: Mixture of Experts with Alternating Policy Optimization for Universal Humanoid Motion 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
Seonghyeon Lim, Hyeonwoo Lee, Mincheol Kim, Kimin Lee, Hyun Myung
MAPO trains a dense mixture-of-experts policy for humanoid motion tracking by measuring router and expert updates separately, holding the other component at its rollout value. It computes both clipped losses from one rollout. In a three-seed simulated teacher-policy evaluation, MAPO reports 87.03 ± 0.64% success versus 82.89 ± 0.43% for PPO with MoE, and 147.9 ± 1.5 mm body-position error versus 154.1 ± 1.7 mm.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.