Abstract
In typical robot safety monitoring, only the violations covered by explicit training or programming are detected. When a new safety rule is introduced, retraining or code changes are needed before the rule can be monitored. This disclosure describes a technique in which a natural-language safety policy document is read at runtime and used to condition a vision-language model (VLM). At each monitoring cycle, camera images and the full policy document are provided to the VLM, and any violation of any policy is identified. Each violation is reported with a reference to the policy, the supporting visual evidence, and a risk assessment along six dimensions. Simultaneous violations are ranked by a fixed order of human, environmental, and robot safety. A safety ontology and zero-shot reasoning are used to flag unsafe scenes that no written policy covers. Detection scope is changed by editing the document, without retraining.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Garikapati, Divya, "Open-Vocabulary Safety Reasoning for Robots Conditioned on Natural-Language Safety Policy Documents", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/12101