ArduPilot's mav-agent lets you command drones with natural language

At the ROS Aerial Robotics community working group meeting on August 26, 2026, ArduPilot dev-team member Sanket Sharma presented mav-agent, an agentic control stack for MAVLink vehicles. The system lets an AI agent perceive and reason over vision, LiDAR, and sensor data plus vehicle state to plan and execute commands via any MCP-compatible client. Users can issue natural-language instructions for movement, visual follow/servoing, obstacle avoidance, and missions with spatio-temporal understanding—no traditional code needed.

Under the hood, mav-agent integrates a MAVLink backend with LangGraph for planning and Qwen-VL for multimodal perception. Sensor data is formatted specifically for agent reasoning, and the MCP tool interface allows external agents or LLM clients to control the drone directly. Sanket demonstrated real-time natural-language control and external-agent handoff. The roadmap targets multi-agent swarm coordination and a spatio-temporal semantic map, promising a significant leap in autonomous aerial robotics accessibility.