Reproduce “PlanDiv: Test-Time Scaling with Latent Plans 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
Hanbo Zhang, Lianlei Shan, Liuhaichen Yang, Zezhi Tang
PlanDiv generates robot action chunks under different latent plans rather than resampling one plan. A planner proposes plans from one backbone pass, and a shared executor decodes a chunk for each. On CALVIN ABC-to-D, five-task success rises from 56.9% with one candidate to 69.9% with eight, versus at most 64.9% for matched resampling; results cover one backbone and robot arm.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.