Collision avoidance
Continuous three-dimensional sensing of the drone's surroundings, so the platform can stay clear of obstacles and other aircraft without relying on a video feed.
Compact AI radar for drones.
Perception that doesn't depend on visibility.
Toro Detection Systems is building the radar sense that small drones are missing. One compact module that perceives the surroundings, recognises what it is tracking and acts on it. In fog, rain, smoke and darkness, and when a target hides against the terrain.
Every drone will need to see for itself. We are building the sensor.
Drones multiplied. Onboard sensing did not.
Small drones are now everywhere, and almost none of them can reliably tell what else is in the air around them.
Cameras and thermal imaging degrade exactly when conditions get hard: fog, rain, smoke, glare, night, or a target sitting against cluttered terrain. Capable airborne radar exists, but at a weight and price built for large aircraft, not for a drone that fits in a backpack.
Hundreds of manufacturers and integrators are building airframes. Very few are building the sensor those airframes need to fly autonomously and safely beyond the pilot's line of sight.
We supply the sensor. We don't compete with the platforms that carry it.
GUARDIAN radar module
One module. Three jobs.
GUARDIAN is a radar perception module sized for FPV-class drones. It processes everything on board and hands the flight controller what it needs to act.
Object classification
Onboard AI reads the micro-Doppler signature of each tracked object and identifies what the radar is tracking, not merely that something is there.
Interception-grade guidance data
Guidance data computed on the module, precise enough to close on a moving aerial target. Terminal guidance is one of the key applications: the radar carries the final approach, when a camera alone cannot be relied on.
Avoid and search, at the same time
Because the module knows what it is looking at, a drone carrying it can do two things at once: keep itself clear of obstacles, and search the airspace for other drones and aerial objects, classifying each one it tracks.
That single capability serves collision avoidance, autonomy, aerial-object detection and counter-UAS applications.
Drops into platforms that already fly
GUARDIAN is interoperable by design. It supports a wide range of flight controllers, with ArduPilot and PX4 supported out of the box.
Radar perception becomes something a drone builder adds to an existing airframe, not a reason to design a new one.
Where this goes
Radar perception as a standard part of every drone.
GUARDIAN is the first step. The company we are building supplies sensing to the whole drone industry.
A sensor built for volume
The production module is being engineered to be small, light and affordable enough to design into drones at scale, including airframes that are not expected to come back.
Autonomy that does not wait for a clear day
Operations are moving beyond the pilot's line of sight. That takes a drone that perceives uncooperative traffic on its own, in any weather, without a video feed to a human.
One core, many missions
Collision avoidance, autonomous flight, aerial-object detection and counter-UAS all draw on the same sensing and classification core. We develop it with the builders who need it.
Classifier research
Classification light enough to run on the drone.
A classifier is only useful on a small drone if it fits the computing budget of a small module. Our models were designed for that budget from the start, and measured against a conventional CNN benchmark on recorded radar data.
33×less computation per classification than the CNN benchmark
+4.3percentage points higher test accuracy, from the same model
| Model | Operations per classification (bars to scale) | Reduction | Test accuracy |
|---|---|---|---|
| CNN benchmarkConventional spectrogram approach | 1× | 88.6%baseline | |
| TDS high-accuracy modelMost accurate | 4.3× | 95.3%+6.7 points | |
| TDS balanced modelBest accuracy for its cost | 33× | 92.9%+4.3 points |
Operations are the approximate number of multiply-accumulate operations each model needs for one classification, the quantity that governs inference time on embedded processors. Accuracy is measured on a held-out test set of recorded radar data with four classes, including drones, people and background noise, for one radar configuration.
Built for real time
A smaller model means a faster answer. The classifier runs on the module itself, alongside detection and tracking, so the drone does not depend on a ground station or a data link to know what it sees.
Multi-target inference in every frame
A small model means fast inference. Fast inference leaves time to classify many tracked objects within a single radar frame, not just one.
Accuracy or economy, by choice
From the most accurate model to one 33 times lighter than the benchmark, the approach scales to the processor and power budget a platform can spare.
How it works
From radar return to a decision, on board.
Spinning propellers leave a fine, repeating modulation on a radar echo, known as a micro-Doppler signature. People, animals and background clutter leave different ones. GUARDIAN uses that difference.
Detect
The radar picks out objects around the drone and measures where they are and how they move.
Track
Detections are followed from frame to frame, so each object keeps its identity over time.
Read the signature
For each tracked object, the module isolates the part of the radar signal that carries its micro-Doppler signature.
Classify
A compact neural network decides what the object is, and the result goes to the flight controller.
Onboard camera
Radar map
Detected objects
Tracked object's signature
Company
Every layer of the stack is ours.
Toro Detection Systems is a deep-tech company based in Warsaw, Poland, working on dual-use radar sensing for unmanned aircraft.
The work grew out of university research on radar signal processing and machine-learning classification of UAVs. We took it from theory to a flying proof of concept, and we are now turning it into a product.
- Radar firmwareRunning on the sensor itself
- Signal processingDetection, tracking and signature extraction
- Edge AI modelsTrained and validated for on-module inference
- Guidance softwareRefined over many flight-test iterations
- Hardware designA purpose-built module in place of evaluation hardware
- ToolingIn-house replay and analysis of synchronised flight data
Work with us.
We are building the company around this technology now. If any of the three below is you, write to us. We share more detail in conversation than we publish here.
For engineers
Build it with us
- RF and PCB electronics
- Embedded firmware
- Machine learning and radar data
For customers and technical partners
Put radar on your drone
- Integration of GUARDIAN into your platform
- Collision avoidance, autonomy and aerial-object detection
- Counter-UAS and other sensing needs
For investors
Back the next step
- Proven concept, flown on real hardware
- Research-backed, in-house technology
- A sensor for a market of many drone builders
Or write directly to mark@torodetectionsystems.com