Reproduce “Back on Track: Deployment-time Policy Improvement via World Model Planning” 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
Colton Crosby, Hao Wang, Mac Schwager, Somil Bansal
The wrapper monitors a fixed policy for stalled progress, briefly steers it using imagined actions, then returns control. It learns a policy-conditioned liveness value from nominal-policy rollouts and uses that signal to guide planning in a world model. In perturbed LIBERO-Goal evaluations, it raises mean success across ten tasks from 72.6% to 81.5%, but interventions can disrupt successful runs.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.