Reproduce “Beyond Validation Loss: Coverage and Reliability Dynamics in Robot Policy Pretraining” 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
Zhanyi Sun, Max Simchowitz, Shuran Song
The authors evaluate robot-policy checkpoints using fixed-start pass-at-k, which measures whether repeated stochastic attempts can succeed from held-out starts, alongside pass-at-1 reliability. They select checkpoints by balancing held-out action-error quantiles and introduce JointAug, which augments demonstrations with generated observation-action pairs. Across studied tasks, selected checkpoints and JointAug improve coverage-related outcomes, but checkpoint rankings are not universal.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.