LLMs Generate Executable Navigation Constraint Functions

Authors propose STPR, a constraint-generation framework that uses LLMs to translate 'what not to do' natural-language instructions into executable Python constraint functions. They show STPR accurately encodes complex mathematical and spatial constraints, integrates with point-cloud planners in Gazebo simulations, ensures full constraint compliance with short runtimes, and works with smaller code-focused LLMs for lower inference cost.