Reproduce “FARM: Reading Failure Signals from the Internal Predictive States of a Frozen Robotic World Model” 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
Haoran Pei, Mingrui Luo, Senbao Wang, Haoran Lv, Jie Guo, Sheng Zhong, Ruixi Ci
FARM uses a small readout to score robot failure risk from frozen predictive-model states, without retraining the predictive backbone. It pools token-level states at each replanning step and tracks the highest score observed so far. Across seven source tasks, it reaches pooled AUROC/AUPRC of 85.68/88.59; strict-Unseen results trail the leading baseline, and intervention is untested.
Key insights
- Failure-relevant information is inherently present and directly decodable within the unaltered internal predictive states of a frozen pre-trained robotic world model.
- A single, general-purpose world model can serve as a reusable feature extractor for complex semantic tasks like failure detection, promoting computational efficiency and reducing the need for new, dedicated monitoring components.
- Lightweight, readout-only adaptation can effectively address domain shifts and improve monitoring performance across different robot platforms, policies, and visual domains with limited target supervision.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.