Anthropic's Project Pilot: Can AI Control a Drone?
Anthropic partnered with Andon Labs to test whether AI models can autonomously control a drone for surveillance tasks, creating the Drone-Bench benchmark with five sub-tasks.
Project Pilot: Can AI Fly a Drone?
In July 2026, Anthropic partnered with Andon Labs to publish Project Pilot research — a series of evaluations assessing AI models' ability to control a quad-rotor drone for a locate-and-follow surveillance task. This extends Anthropic's previous work on AI-physical world interaction, following Project Vend (AI running a shop) and Project Fetch (robot dog retrieving objects).
Drone-Bench: Five Sub-Tasks
Andon Labs deconstructed the core locate-and-follow task into five sub-tasks: Reconstruct (turning videos into a 3D model and obstacle map), Localize (matching drone's current view to the 2D map), Navigate (planning a path between rooms), Detect (finding a target person from a reference photo), and Follow (keeping the target centered in view as they move).
What This Means
The research demonstrates that frontier AI models can chain complex hardware-control tasks together. Anthropic's Frontier Red Team conducts such evaluations to maintain situational awareness about how close AI is to autonomous hardware control — with all the benefits and risks that entails, from agriculture and search-and-rescue to surveillance and dual-use concerns.