Free Drone Software Lets It Fly Blind Through Unknown Buildings
No map, no pre-set route — the drone figures out the building as it flies.
A developer named formiat published an open-source drone navigation system on GitHub. The idea is simple to describe and hard to build: give a drone a starting point and a destination inside a multi-storey city, hand it no map at all, and let it work out the way there. As it flies, the drone sweeps the world with a simulated 3D lidar (a laser sensor that measures distance to objects) and paints everything it sees into its own growing three-dimensional memory of walls, floors and open space.
On top of that memory sits the route-finding brain. A "planner" called D* Lite keeps a running plan and quickly patches it when a new wall appears, rather than starting over from scratch. A second piece of software, running on an NVIDIA graphics card, checks 50 times a second that the drone's next fraction of a second of movement is physically safe. All of it runs through PX4, the free autopilot software already used by many drone makers. The test world is a real one: the urban course built for DARPA's Subterranean Challenge, a US government robotics competition set in mock disaster sites.
The published results are modest but clean. Five flights back-to-back on the same code, no changes in between, each covering roughly 370 to 417 metres in about two to two and a half minutes at around 3 metres per second — a brisk jogging pace. Zero crashes. When an obstacle surprised it, the planner produced a new route in about 0.15 seconds, and the drone braked while it waited.
The most interesting design choice is also the riskiest. The drone treats space it has never looked at as safe to fly through, rather than stopping to peek around corners. What makes that defensible is one rule: it never flies faster than it could stop within the distance its sensor is guaranteed to have cleared. That works out to about 5.6 metres per second in level flight, and it flew slower than that. The catch is important: this is simulation only, by one developer, and explicitly not validated for real aircraft.
- It's free and open source (MIT licensed), so anyone can download, study or build on the code
- In five simulated flights of about 400 metres each at jogging speed, the drone crashed zero times
- The drone runs no map at all — it remembers obstacles as it goes and re-plans in about 0.15 seconds when surprised
Why It Matters
Cheaper, map-free drone navigation could speed up search-and-rescue, inspections and deliveries in places nobody has surveyed.