Veg, a drone that flies without GPS

Veg means speed in Sanskrit. It is our research quadcopter for places where satellite positioning fails: inside buildings, under cover and in cluttered industrial sites.

Now: building Veg 3.0 in simulation

How Veg works

Veg runs four jobs at once, all onboard. A camera and an IMU feed ORB-SLAM3, which tracks the drone in six degrees of freedom and closes loops when it returns to a place it has seen. The map points become an occupancy grid, and a Dijkstra planner finds a route through it.

An Arduino Nano holds the drone steady with an LQR attitude controller while the Raspberry Pi 4 runs a position controller on top of the SLAM pose. A fault monitor watches the motors and the attitude response. When a rotor fails, Veg picks the nearest safe landing zone on its map and descends there.

A lightweight CNN detects people and vehicles, and a PCA model recognises known faces. Both run beside the control loop, never inside it, so heavy vision work cannot make the drone unstable.

Veg system diagram: IMU and camera feed ORB-SLAM3, mapping, object detection and fault detection, which drive the quadcopter
Veg system overview.

Results from the paper

Simulation and lab results for the research version, as reported in our arXiv preprint.

Attitude control
LQR settles a 5° roll step in 0.1 s, against 3 s for the PID baseline
Maze flight
5 of 5 simulated runs reached the goal, average deviation under 0.35 m
Rotor fault
Detected in 1.3 s on average, safe landing in 8 of 10 simulated runs
Object detection
About 2 frames per second on the Pi CPU at 91.7% precision
Face recognition
94.6% accuracy on a 20 person test set
Electrical draw
About 4.3 W for SLAM, vision, fault monitoring and control together

What comes next

Veg 3.0 is being built simulation first. We are building a full simulation of a SLAM based drone, and every part of the system is tested there, with faults injected on purpose, before it is trusted in the air. Hardware comes after the simulation proves the design.

If you want to follow the work, partner on a pilot or join the team, write to us.

Maze map showing the original mission path, the rerouted path after a rotor fault, and the fault location
Rerouting to a safe landing zone after a rotor fault, in simulation.

Reading on Veg and aerial robotics