PEACE: A Planner-Executor Agent with Constraint Enforcement for UAVs

Researchers propose PEACE, a planner-executor agent architecture for autonomous drones that decouples high-level mission planning from low-level control using foundation models. The system combines large language models for task planning with structured tool-calling interfaces and constraint enforcement mechanisms, demonstrating improved explainability and reduced computational overhead compared to tightly coupled LLM approaches.

PEACE addresses a fundamental challenge in foundation-model-based robotics: the tension between leveraging large language models' reasoning capabilities and maintaining system safety, interpretability, and efficiency. The research demonstrates that decoupling planning from execution—rather than keeping LLMs in tight control loops—reduces hallucination risks and latency while enabling explicit constraint enforcement critical for safety-sensitive applications like autonomous aerial vehicles.

This work reflects broader industry recognition that end-to-end neural approaches, while powerful, sacrifice transparency and controllability. The layered architecture combining modular vision components (YOLO, vision-language models) with depth projection for 3D localization and explicit geofencing rules follows established robotics principles enhanced by LLM capabilities. The use of ROS 2 and MAVLink standards ensures practical integration with existing drone ecosystems rather than requiring proprietary implementations.

The significance extends beyond UAV research to the general pattern of how autonomous systems should incorporate foundation models. By positioning PEACE within established design patterns and validating it through simulation, the authors provide a replicable blueprint applicable to ground robots, maritime vessels, and other domains where safety constraints are non-negotiable. The promise of open-sourced code and datasets accelerates community adoption.

Key technical innovations include single-pass task planning to minimize LLM invocations and bounded replanning mechanisms for runtime failure recovery. These features directly address production concerns: computational costs and graceful degradation. The constraint enforcement layer exemplifies how rule-based systems complement rather than compete with neural approaches.

• →PEACE decouples LLM-based planning from low-level control, reducing latency and hallucination risks compared to tightly coupled architectures.
• →The system implements explicit constraint enforcement for altitude limits and geofencing, critical for safety-critical autonomous systems.
• →Integration with PX4, ROS 2, and MAVLink demonstrates practical compatibility with established drone software ecosystems.
• →Modular vision components combined with depth projection enable 3D object localization without domain-specific fine-tuning.
• →Open-source release and validation in Gazebo simulations provide reproducible foundations for broader robotics research.