Reproduce “SWAP: Stepwise Action Policy Routing for Vision-Language-Action 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

Mousumi Das, Aditeya Prajapati, Abrar Anwar, Jesse Thomason

SWAP trains an offline router to select among fixed robot controllers at action-chunk boundaries, using current observations and predicted long-horizon return. It learns from successful and failed rollouts, then reevaluates after each chunk. On DROID, it averages 96.7% success across three tasks, versus 63.3% for the best fixed policy; in shifted LIBERO-Plus testing, its result falls when policy strengths change.

Summary by alphaXiv.

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