< CitiNavAgent: Zero-Shot Drone Navigation in Cities >

"Fly to the white statue after passing the red phone booth" is now a fully acceptable command thanks to CitiNavAgent

The authors design a VLN (visual-language-navigation) model that uses a hierarchical semantic planner to break long-horizon instructions into subgoals of varying abstraction, and leverages a global memory graph of past trajectories to simplify navigation in familiar areas

Let's break it down: arxiv.org/abs/2505.05622

🧵1/8

< Problem >

Aerial visual language navigation is really complicated

> No prebuilt graphs
> Continuous 3D action space
> Sparse semantics at altitude
> Longer, ambiguous instructions
> No guarantee of seeing your goal

🧵2/8

< Aerial Edge AI >

From sensors, to chips, to decisions. How is the drones brain wired?

Today, we're tracing the data lines that power AI & control onboard the flying machine

🧵 1 / 7 The drone processes a ton of raw data in real time.

Camera, IMU, Barometer, GPS...

But how does that flow into decisions?

🧵 2 / 7

< Fundamentals of AI in Drones >

Dissecting the fundamentals of drone state estimation and sensor fusion in quadcopters

We're talking about sensors at high refresh rates, noise filtering with probabilistic filters, and predictive model controls that tie it all together

🧵 1/7 The Sensor Suite

Camera (@ ~30–60 FPS) - captures visual data for object detection, tracking, and navigation

IMU (@ ~100–200 Hz) - measures acceleration and rotational rates for fast attitude correction

GPS (@ ~1–10 Hz) - global positioning, often noisy and with respective tolerance

Barometer (variable refresh): assists in altitude estimation

These form the backbone of aerial state estimation

🧵 2/7