Reproduce “Beyond Policy Alignment: Closing the Planning-Learning Loop for Robot Control with Learned 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

Kowndinya Boyalakuntla, Yuhan Liu, Abdeslam Boularias

PL-MPC changes how planner experience trains the critic and policy and how the planner scores terminal values. It uses hybrid multi-step targets, discounts terminal values when target critics disagree, and weights imitation of planner actions by episode return. On HumanoidBench balance-hard, mean return rises from 98 ± 18 to 387 ± 255 versus TD-M(PC)², with substantial seed variability; gains remain task dependent.

Summary by alphaXiv.

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