Live tests // FMV Battleweek // Älvdalen, 16–18 Sept 2026
Three days live on a real sensor tower.
FPV drones on Wednesday. Ground vehicles on Thursday. A Shahed-type fixed-wing drone on Friday. A sound radar, RGB and IR cameras and sensor fusion went in. A short list of alerts, each with a reason, came out on one operator screen, with every decision recorded.
All numbers come from MIMIR's own tamper-evident log · 17 939 rows, chain intact
These live tests were FMV Battleweek.
Battleweek is run by FMV Test & Evaluation, the test arm of the Swedish Defence Materiel Administration. It is a recurring, structured way to find, test and demonstrate new solutions that may matter to the Swedish Armed Forces and to Ukraine. FMV puts military challenges to the market; selected companies, mostly small and medium-sized, bring their solutions to FMV's test range to show them working.
In September 2026 that range was the airstrip outside Älvdalen. MIMIR ran there for three days on the Askr sensor tower, against real targets, with the software changed on site whenever the screen showed us something to fix.
FMV hosted the week. Taking part is not an evaluation result. About Battleweek at FMV ↗
- Wed 16Small FPV droneswith cars, a tent and people in the areaCan the system tell a drone from normal ground activity?
- Thu 17Ground vehiclescars and trucks, parked and driving, and peopleWhat is it, is it moving, is it coming closer?
- Fri 18Shahed-type fixed-wing dronemultirotor drones in the morningIs a fast drone at long distance reported early enough?
The sensors see and hear. MIMIR decides what a person must see.
Askr sensor tower: RGB and IR cameras on a gimbal
The tower turns its cameras towards what the sensors report. Several times a second it sends MIMIR a list of tracks. Each track is a direction: a compass bearing and a height angle. The tower measures no distance. With each picture it also says where the camera points and how far it is zoomed in.
Sound radar: the direction of a sound
An acoustic array that hears an engine and gives its direction, often before any camera has it. A sound alone is a hint, not a contact: voices, cars and wind are sounds too. On Friday it was the sound radar that followed the big drone's zig-zag first.
Laplacian: sensor fusion
Joins the sound radar and the cameras into tracks and sends them to MIMIR over a signed, labelled wire (STANAG 4774/4778 checked on every frame).
MIMIR: the operator's screen, and the judgement behind it
- The complete operator interface: plan view, bearing view, camera view, alert list, object history and chat. One to three screens, Swedish or English.
- It decides what gets priority. A sound is a hint. A camera-confirmed object is a contact. Something coming towards the tower outranks something passing by.
- It analyses the camera pictures itself with its own vision model (NVIDIA Nemotron) and says what the object is.
- It analyses the situation and talks to the operator through the chat Agent M, with its own language model (Gemma 4).
- It learns from the operator, storing every decision next to its own judgement, and records everything in a log where any later change can be detected.
- Both models ran on one NVIDIA Jetson Thor beside the tower. No cloud. Nothing left the site. MIMIR has no connection to any weapon.
What happened, and what we changed while it happened.
The software was changed during each day, from what we saw on the screen: 32 updates were loaded at the mast in three days. So these figures describe what happened. They are not a measurement of a finished system, and we claim no detection rate.
Question Can the system tell a drone from normal ground activity?
MIMIR is up and the tower feed is live. First alert at 08:47.
Every track arrived with the same air label, "aircraft, unmanned". MIMIR believed the label, so parked cars on the airstrip became air alerts at high priority.
Software update at the mast: MIMIR sets the label aside, analyses the camera picture with its own vision model and says what the object is. Alerts at the two highest priorities fell from 46 to 25 for the same length of time.
The first version called two parked cars and a tent "a small white aircraft on the runway". Corrected: only an answer that says "this object is on the ground" can lower an air alert.
An older test camera path is switched off, so every picture judged is the tower's own.
2.6 hours live. 791 pictures analysed in 80 minutes.
Four updates between 20:46 and 23:20: an event journal that exports as one file, vehicle type and motion from angles alone, an early warning from sound, "air contact coming closer" with time to arrival, and joining the many track numbers one object can get. During the night MIMIR wrote its first profile of the operator from the day's 36 decisions.
Before and after 10:06, drag to compare
What the 791 pictures showed
Question What is it, is it moving, is it coming closer?
The operator enters the position of the mast. First pictures at 08:08.
Movement words must match what was measured. A building is never the tracked object while something that can move is in view. One alert per direction (one car had got six). The vehicle type is a vote of the pictures, named only when more than half agree.
Person, vehicle and drone labels arrive with the tracks. MIMIR's vehicle alerts become immediate.
Parked, passing, leaving, approaching. MIMIR names the type and says what the vehicle is doing, with an estimated time to arrival.
No sound warning on a tracked ground object. An alert at the same level changes its text at most every 8 seconds. A jump of more than 8° between two readings is a new object.
Each track is marked in the picture and the object at that point is judged. 213 pictures were judged this way the same afternoon. One operator language across the whole screen. A sound warning ends when the camera has the object.
6.5 hours live. MIMIR was restarted 19 times this day to load fixes, each from something seen on the screen.
From angles alone
The tower measures no distance. MIMIR uses how fast the angle down to a vehicle changes: closing traffic drops in the picture faster and faster, so a time to arrival can be estimated without knowing the range, and the height of the mast cancels out. The airstrip rises away from the tower; we measured the horizon on site (+0.6°) so "far away" is not mistaken for "coming closer".
Vehicle types named (pictures)
What the 494 alerts were about
Question Is a fast drone at long distance reported early enough?
Multirotor drones in the morning.
People talking next to the mast. MIMIR warned of an airborne sound source 9 times in 18 minutes.
A person the camera sees in the same direction ends the warning. The same recording then gives 1 warning instead of 9. MIMIR also reports whether a person wears camouflage or civilian clothes.
A drone can no longer be called a tent or a ground vehicle, and a track that climbs into the sky cannot keep the label "car".
Thousands of camera detections that each lived under a second. MIMIR's vision model saw nothing in any of the 4 020 pictures it checked, yet 183 alerts were raised on them, all of which expired. Our lesson: a track that lives only a moment must not reach the reasoning model.
Every run simulated against the live code. Straight runs raised "air contact coming closer"; zig-zag runs raised nothing.
Altitudes 50 to 600 metres, 35 and 50 m/s, straight runs from 5 km, zig-zag runs from 2.2 km, circles at 1.5 km.
Fix loaded eight minutes after take-off. In the simulation every run at 50, 100, 300 and 600 metres now raises the alert.
The operator enters the flight altitude (50 m); MIMIR then shows a distance, marked as calculated, not measured. A large drone is never called a manned aircraft in an alert.
For 105 seconds the sound radar followed the drone weaving towards the tower; no camera had it for 104 of them. Replay it below.
5.8 hours live. The fixed-wing flight, about 55 minutes, recorded in full.
The day in three parts, select one
Multirotor drones, 07:33–11:10. 439 alerts · 293 decisions (244 approved, 47 rejected, 2 escalated).
Sensor noise, 11:10–12:25. No operator activity. 183 alerts on detections that lived under a second. All expired.
The fixed-wing drone, 12:25–13:21. 98 alerts · 66 decisions (50 approved, 15 rejected, 1 escalated). 43 of the alerts: "air contact coming closer".
Alerts by highest priority reached: Friday
| Wed | Thu | Fri | Total | |
|---|---|---|---|---|
| Live time (hours) | 2.6 | 6.5 | 5.8 | 14.9 |
| Alerts shown to the operator | 127 | 494 | 720 | 1 341 |
| Operator decisions | 36 | 49 | 359 | 444 |
| of which approved / rejected / escalated | 27 / 0 / 9 | 36 / 13 / 0 | 294 / 62 / 3 | 357 / 75 / 12 |
| Camera pictures analysed by MIMIR | 791 | 5 253 | 7 178 | 13 222 |
| Time per picture, median (s) | 2.7 | 1.8 | 0.9 | |
| Software updates loaded on site | 3 | 19 | 10 | 32 |
The drone flew for about 55 minutes. MIMIR recorded the whole flight.
The zig-zag was heard before it was seen.
For 105 seconds the sound radar followed one source weaving towards the tower. For 104 of those seconds no camera saw it. The operator could read the flight pattern straight off MIMIR's bearing view. Press play to replay the real data.
Bearing → · time ↓ · the six turns marked
- 12:56:05104°
- 12:56:2080°
- 12:56:35100°
- 12:56:4780°
- 12:57:02100°
- 12:57:1380°
What MIMIR did with it
The track was on the screen, but MIMIR raised no alert of its own for it: its sound-warning rule expects a steady direction, and a weave is not steady. The same blind spot showed up earlier in software. That is why the camera-based "coming closer" rule was fixed at 12:33. The sound rule is next. Still open.
Seen on the screen. Live at the mast the same day.
A system that records every input can be replayed, and a system that can be replayed can be corrected in minutes. Each marker is a problem we saw and the moment its fix went live.
444 decisions, each stored next to MIMIR's own judgement.
Every click (approve, reject, escalate) was saved with what MIMIR had judged at that moment, the text the operator saw, the sensor data and the age of the picture when the operator clicked. The learning ran during the tests, not after them: each night MIMIR wrote a profile of how consistently the operator decides. Twelve escalations each produced an incident report automatically.
When MIMIR's judgement matches the operator's for one type of object, it can suggest handling that type by itself. The customer decides if that is allowed. People are never filtered out.
About five weeks of an operator working the console gives MIMIR enough judgements to propose its first class for autonomous handling.
The 444 decisions
All 444 decisions come from one operator. Measuring how far different operators agree with each other needs more than one. That is the next step, not a result.
Decide, then see what the operator did, and what MIMIR keeps.
Six alerts exactly as MIMIR showed them at Älvdalen. There is no right answer here: the point is that every decision, yours or the operator's, becomes a line MIMIR learns from.
What the three days told us to build next.
- Open
Ignore what lives only a moment
Sensor noise can produce thousands of detections that last under a second. A track that short, which the picture analysis cannot see, must never reach the reasoning model.
- Open
Warn on a weave heard by sound
A regular zig-zag from sound alone is on the screen but raises no early warning yet.
- Open
Understand the vehicle rejections
On Friday the operator rejected 22 of 26 vehicle alerts. That disagreement is exactly what the corpus is for. It gets examined, not hidden.
- Open
A drone that circles
A drone holding a constant distance never "comes closer", so the fixed rules raise nothing for it.
- Open
A real distance
Without a range sensor MIMIR shows a distance only when the operator declares the flight altitude. A rangefinder or a second tower would give a measured one.
The whole chain works with real sensors.
MIMIR takes soldiers off the watch, not off the trigger.
Request the full report
Askr sensor tower and Laplacian sensor fusion: Xymbiotec · Sound radar: Monava · Hosted by FMV Test & Evaluation
All numbers come from MIMIR's own logs · No detection rate is claimed