Show HN: URML – safety-eval harness for AI agents on lab and factory hardware (github.com)

URML, a newly introduced language for defining robot intent, aims to enhance safety evaluations for AI agents operating in lab and factory environments. This framework focuses on validating whether an agent's proposed actions align with pre-defined hardware limits and operational envelopes, providing machine-readable reasons for any rejections. While it doesn't assess physical capabilities directly, URML emphasizes the accountability of integrators and vendors in ensuring the authenticity of declarations related to hardware constraints. Its architecture allows for validating intents against declared limits, ensuring that the strictest conditions guide the decision-making process. The significance of URML lies in its potential to advance safety standards in AI applications for physical systems, especially as the AI community prepares for standardized protocols such as the Model Hardware Standard (MHS). It supports a methodical approach to safety evaluations, rehearsing each proposed action under predetermined motion models and logging rejections along with evidence tags for further insights. By integrating URML with a scaffold for read/write calls through a defined transport system, this tool sets a foundation for reliable automation in complex environments, ultimately boosting confidence in deploying AI agents in critical tasks without compromising safety.