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

Abstract:

While foundation models are gaining traction in autonomous systems, current methodologies present significant challenges. Approaches that maintain models within tight control loops often suffer from increased latency and higher risks of hallucination. Alternatively, methods that translate natural language into opaque, end-to-end policies lack explainability, necessitate domain-specific datasets, and require extensive fine-tuning.

To address these issues, we introduce PEACE, a planner-executor agent designed for PX4-based drones. This architecture separates high-level mission planning from low-level control mechanisms. A large language model (LLM) executes single-pass task planning, while a structured ROS 2 tool-calling interface, bridged to MAVLink, manages the execution phase. The system builds a comprehensive world model by integrating modular 2D detection systems, such as YOLO or vision-language models, with a pinhole depth projection module to achieve 3D object localization.

Furthermore, the system incorporates a constraint enforcement layer to ensure compliance with altitude limits and horizontal geofencing. It also features bounded replanning capabilities to recover from action failures occurring during execution. We contextualize our approach among three prevalent design patterns for foundation-model-based robotics and validate its feasibility through PX4 software-in-the-loop simulations in Gazebo. Our results indicate that this method enhances explainability and constraint adherence while reducing the frequency of LLM calls compared to tightly coupled LLM control strategies.

The codebase, dataset, videos, and supplementary materials are available at: https://github.com/erdemuysalx/PEACE

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC