Reproduce “No Free Checker: A Survey of Verifiers for Robot Policies” 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
Yang Wan, Xihang Yue, Zhirui Liu, Ziyuan Chu, Shuxun Wang, Yuhan Chen, Xiaonan Jiang, Xukun Zhu and 2 more
“No Free Checker” surveys roughly 150 verification methods for robot policies, spanning human judgments, formal rules, learned evaluators, and model-internal signals. It organizes them around an availability–credibility trade-off and proposes nine metrics for assessing whether a verifier remains reliable across real, generated, optimized, and distribution-shifted robot behavior.
Key insights
- Verifier availability and credibility generally trade off: dense, inexpensive signals such as intrinsic confidence or learned rewards are easier to query but require stronger assumptions and validation than human or explicit formal judgments.
- Verifier errors are use-dependent. A false positive may contaminate demonstrations, misdirect candidate selection, or become a reward-hacking target, while runtime systems additionally face latency and safety consequences.
- Performance on ordinary held-out rollouts is insufficient evidence of robustness, because policy optimization changes the data distribution and can actively search for behaviors that score well without satisfying the intended task.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.