betaFeedback

WorksEngineering

NL2RCtrl: Risk-Controlled Controller Synthesis from Natural Language Requirements

Abstract

Large language models (LLMs) have enabled the translation of natural language (NL) requirements into linear temporal logic (LTL) specifications, facilitating controller synthesis for cyber-physical systems (CPSs). However, the reliability of such translations remains a concern in safety-critical applications, particularly when NL requirements are ambiguous. This paper proposes a framework based on Conformal Risk Control (CRC) to provide a statistical guarantee on the risk of issuing a controller that violates the user's intended requirement. Rather than requiring semantic equivalence between the generated and intended specifications, the proposed approach considers language inclusion, a sufficient condition for ensuring that the traces of a controller synthesized for the generated specification satisfy the intended requirement. To account for ambiguity, the LLM is queried multiple times, and the resulting LTL specifications are grouped into semantic equivalence classes and ranked by their frequency of occurrence. Given a prescribed risk level, CRC selects how many of the top-ranked equivalence classes to combine into a single LTL specification, such that the probability of issuing a controller that violates the intended requirement is bounded by that level. A controller is then synthesized for the resulting specification, or the framework abstains. Experiments on a vehicle navigation case study evaluate the proposed approach.

Paper details

vlnc:W_4tatwrsu
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
arXiv nonexclusive license
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Engineering

Generative AI tools

View statement
Looking for related papers…