THORONDOR · AUTONOMOUS SYSTEMS

Thorondor.

Thorondor builds autonomous drone systems, with the software designed in France and fully under our control. Two products share the same perception core. Chronos: an autonomous reconnaissance drone that surveys an area on its own, detects people and vehicles, places each of them on a map, and reports what it found, with all perception running onboard. Talos: the other side of the same problem, protecting a site against small drones by detecting, tracking and intercepting them.

Why

Small drones are now everywhere, but most still need a pilot watching a video feed. Thorondor focuses on the layer that removes that bottleneck: onboard perception and autonomy. The drone turns raw camera images into a clean, de-duplicated map of what is on the ground, with an honest estimate of how certain it is, so that an operator gets information instead of footage.

Applications include search and rescue, site security, infrastructure inspection and situational awareness.

What already works

The full perception pipeline runs today in simulation (PX4 autopilot and Gazebo), on a quadcopter flying autonomously.

Demo, in simulation (67 s). Left: the drone's camera with detections. Right: the map the drone builds live. First an autonomous area survey, then a walking person followed as one single tracked object. Near the end of the survey, one vehicle (c2) drops off the map: see the explanation below the second map.
~20 ms real-time object detection per frame on GPU
0.14 m median error when placing a detected object on the ground
7 / 7 objects found during an autonomous area survey (4 people, 3 vehicles)
Top-down map of the simulated area: the drone's lawnmower flight path, and every detected person and vehicle placed next to its true position.
Autonomous area survey, in simulation. The drone flies a lawnmower pattern at 5 m (red line) and builds this map live. Orange dots are people, blue dots are vehicles, crosses are the true positions; circles show the remaining position uncertainty. Result: 4 / 4 people and 3 / 3 vehicles found, median position error 0.14 m.
Top-down map of a survey flown at 12 m with a gimbal camera: four people and two of three vehicles placed next to their true positions; one vehicle, c2, is missing.
Same survey at 12 m, with a stabilised gimbal camera. People are now seen from almost straight above. Result: 4 / 4 people, 2 / 3 vehicles, median position error 0.14 m, one duplicate (vehicle c3 counted twice).
Why vehicle c2 disappears. c2 is an SUV model deliberately left out of the detector's training set, to test how it copes with something new. Seen from the side on the first passes, it is detected and confirmed on the map. On the last pass, it is seen from above and in front, fully in view, and the detector no longer recognises it: the map sees an area where c2 should be visible but nothing is detected, second after second, and removes it. That is the intended behaviour: the map prefers forgetting an object to inventing one. Never removing a confirmed object would have kept c2, but also left ten false alarms on this flight. The real fix is on the detector side: more variety in training, and real images. The detector used here was fine-tuned on simulated images to show the full pipeline; real flights need a detector trained on real aerial footage (VisDrone, then images from our own drone).
Top-down map: the estimated path of a walking person (orange) follows the true rectangular path; the true position is a magenta cross.
Following a walking person. The drone hovers at 12 m, camera pointing down, while a person walks a 8 × 10 m rectangle. The orange line is the path the drone estimated; the magenta cross is the true position. A Kalman filter estimates position and velocity, so the person stays one single object, turns included. Median error 0.31 m.
  • Flight
    Autonomous mission

    Take-off, stable hover, scan and landing, fully scripted.

  • Perception
    Detection and geolocation

    Every detection is projected from the camera image to a position on the ground, for any position and attitude of the drone.

  • Mapping
    Object registry

    Repeated sightings of the same object are fused into a single entry with an uncertainty estimate; false alarms fade out when the drone looks again and sees nothing.

  • Tracking
    Moving targets

    Moving objects are tracked with their velocity: one identity per person, predicted through short occlusions, while parked vehicles keep the precision of a static model.

Talos: detect, track, intercept

Small drones are cheap, fast and hard to see. Talos, our counter-drone system, protects a site against them: it spots incoming drones from far away, follows each one, and sends an interceptor drone to stop it. The whole chain runs autonomously in simulation today; the ground truth is only used to score the result.

Demo, in simulation (30 s). Two drones approach the site at the same time: one at 20 m/s, one at 50 m/s, faster than our interceptors. Top left: a ground camera, where each drone is only a few pixels wide. Top right: the turret camera, locked on one target. Bottom left: top-down map of the tracked drones and of the two interceptors (red). Both targets are hit.
cm-level 3D position of drones tracked several hundred metres away
2 / 2 targets intercepted, including one faster than the interceptor
0 human input between first detection and impact
  • Detect
    See small drones early

    Passive ground cameras watch the approach and pick up drones that are only a few pixels wide.

  • Track
    Know where each drone is, and where it is going

    Every drone becomes one track with an accurate 3D position and velocity.

  • Observe
    Turret camera

    A pan-tilt camera stays locked on a track and gives the close-up view needed to identify the target.

  • Intercept
    Autonomous interceptor drones

    One interceptor per target, launched automatically and guided until contact, including against targets that change direction.

Next: faster targets, more of them, and field tests with our own drones on an authorised site.

Roadmap

  • Next
    Keeping the target in view

    Steering the drone and gimbal so a tracked person stays in frame, and time-stamping the camera attitude so positions stay accurate during fast turns.

  • Then
    Detector for real footage

    The simulation detector only proves the pipeline. For real flights: a detector trained on real aerial datasets, then refined on images from our own drone.

  • Then
    Physical prototype

    Quadcopter with an onboard computer and a stabilised gimbal camera, running the same software as in simulation.

  • Then
    Field tests

    Real flights on an open test site, measuring detection and positioning accuracy outdoors.

  • Later
    Learned navigation

    Flying without GPS and through obstacles, with navigation policies trained in simulation.

Support the project

Thorondor is moving from simulation to real prototypes. If you can help with equipment, a test site, advice or an introduction, see what the project needs.

Needs & partnerships → Get in touch ↗