What Patrol is
An educational simulation on synthetic data. It shows how sensors, drones, a fixed wing and ground teams share the work on a fictional border, how the AI proposes an assignment and how the operator decides.
Simulation, synthetic data
ARES: a fictional system
ARES is the name of a fictional system in this simulation. The product is called Patrol. Several real companies use the name ARES, some of them in defence. Our fictional ARES is not a product of any of them and does not copy any real system.
What it is
The terrain, the line, the fence, the towers and the contacts come from a seed. State A and State B are made up. Every run can be repeated bit for bit on any computer.
What it does not do
- It is not an operational system and it is not for making decisions about people.
- It does not identify anyone: the figures are anonymous, with no faces and no biometrics.
- It does not connect to real hardware: no flight plan export, no autopilot protocols.
- It contains no use of force and no jamming.
Asset classes (fictional names, parameters [A])
- TL-Q4, mini multirotor: 4.2 kg, 199 Wh pack (2 × 99.5 Wh), from the model about 24 min hovering and 28 min cruising, up to 18 m/s, 70 to 90 m above ground, EO and IR camera 640 × 512.
- TL-F12, fixed wing with vertical take-off: 14 kg, 1,600 Wh pack, cruise 20 m/s, stall 13 m/s, 250 m orbit, 400 m above ground.
- TL-M20, 22 m mast with a PTZ EO and IR camera (IFOV 0.12 mrad), slew 30°/s.
- TL-V4, ground team with a vehicle: 12 m/s on roads, 4 m/s over terrain.
- TL-SAT, observation satellite: 95 min period, 15 km swath, 0.5 m pixel, image after 30 to 90 min, never a live picture; clouds block the view.
Fictional name. Parameters: an assumption, a range from public literature.
AI proposals and the operator's approval
- The default mode is proposals: every 5 s the planner proposes a dispatch, a change of task, an escalation or a search, with the arrival time, cost, battery share and risk.
- Nothing happens without the operator's approval except the patrol. Exceptions: return at the battery reserve, return after a lost link, first aid by a team already on the spot.
- A rejected pair of asset and contact does not return for 120 s. The window measures the mean decision time.
Sensors, link and satellite
- Cameras: Johnson criteria (detection, orientation, recognition of the class); there is no identification of persons.
- The ground radar misses an object whose radial speed is below 0.3 m/s.
- Seismic sensors (person 30 m, vehicle 60 m) and vibration sensors on 16 fence segments also give false alarms from animals and wind.
- The radio link depends on the line of sight to the relay over the terrain; after 30 s without a link the asset returns by itself.
Rules of engagement in the simulation
- Groups of people: observe, assess, give first aid when someone needs it, and hand over to the services. Assessment only in State A.
- Armed contact in State B: observe from our side at 400 m or more and report to the services. Ground teams stay away.
- Infrastructure alarm: a patrol checks the spot, a team repairs the segment, the sensor is tested.
- Our drones and teams never enter State B.
What the model computes
- Terrain, border, forest and the movement of contacts come from the seed in the page address. The same seed gives the same run on every computer: the Rust core runs in WebAssembly without functions that differ between browsers.
- Detection is a draw, not image recognition. The chance grows with the pixels on the object under the Johnson criteria (drone camera 640 × 512 pixels, field 36 × 29°) and falls in fog, rain and forest. Recognition covers the class only: person, vehicle, animal, unknown air object.
- A contact track joins the sensor readings. The class probability rises with image quality, and the position uncertainty grows while nobody looks.
- The multirotor uses its battery under the power model of Zeng, Xu and Zhang (2019) for a mass of 4.2 kg and a 199 Wh pack (2 × 99.5 Wh; from the model: about 24 min hovering, about 28 min cruising at 10 m/s), has limited acceleration and speed, and returns for a pack swap with a 25% reserve. The fixed wing cannot hover: it circles on a 250 m orbit and turns no tighter than a 30° bank allows. The teams drive at 12 m/s on roads and 4 m/s over terrain.
- A group walks at the speed Tobler's function gives for the slope. A team on scene assesses the situation, gives first aid when someone needs it, and hands the persons over to the services.
- Every 5 s the planner computes an assignment with the Hungarian method or the nearest free asset rule. In the proposal mode it stores the result as proposals and waits for your decision; in the automatic mode it dispatches by itself.
- The run slider rebuilds the simulation from the seed and the log of your commands, step by step, so a replay gives exactly the same result.
Parameters marked [A] are assumptions within ranges from public literature. They show differences; they do not rate real equipment.
In-browser language model (optional)
- Qwen3.5 0.8B (Qwen Team), Apache-2.0 licence (card: https://huggingface.co/Qwen/Qwen3.5-0.8B). ONNX q4f16 weights from onnx-community/Qwen3.5-0.8B-Text-ONNX, revision 25cdf33f386b487a25f3d743e86442517bdb77ff (https://huggingface.co/onnx-community/Qwen3.5-0.8B-Text-ONNX; that card has no licence field of its own, the licence comes from the base model). Size: 489 MB, downloaded only after a click.
- It downloads only after a click, runs on this computer's graphics card (WebGPU) and sees only the narrator's latest reports. Every model sentence carries the label "generated by a model, may be wrong"; an answer that names an object outside the log is rejected. On a phone and without WebGPU the model is unavailable.
- Libraries: Transformers.js 4.3.1 (Apache-2.0) and ONNX Runtime Web 1.31.0-dev.20260914-8d85527a0 (MIT), served from this site. Nothing is sent anywhere.
3D models
Every shape (a 1.1 m multirotor, a fixed wing with a 3.5 m span, a 4.8 m vehicle, a 22 m mast, figures, animals) is procedural: built in the window's code (file models.ts) at true scale, with no model files to download (0 B) and no copy of real equipment, so no model file licence applies. three.js (MIT) draws them in the scene bundle: 140 kB after brotli compression. Revision: branch r3-scene, commit d05fbc0.
Sources and licences
- Hungarian method: H. W. Kuhn, Naval Research Logistics Quarterly 2 (1955), pp. 83 to 97.
- Walking speed: Tobler's hiking function (1993), approximated without the exponential, never faster than the original.
- Multirotor energy use: the rotorcraft power model of Zeng, Xu and Zhang (2019), TL-Q4 class parameters as an assumption.
- Random numbers: PCG32, M. E. O'Neill (2014).
- 3D graphics: three.js (MIT licence). Geist and Geist Mono typefaces (SIL Open Font License 1.1).
- The full list of assumptions, sources and licences: tenzanlogic.com/en/data-sources/