ALAYA · TACTICAL LAB
Turn limited performance into effective combat power.
ALAYA is designed for small units and base-defence teams to work out the most effective division of labour for the drones they already have: compared in simulation first, then verified on real aircraft.
- ALAYA
- multi-drone mission planning and simulation for small units
- ZERO
- defence language model, deployable on-premise: turns requirements into computable conditions and explains the results
- BABYLON Market
- selecting and assembling vehicles and components
Watch. Try. Verify.
Start here
Tap a picture to open it. Tiles marked “Research account” or “Separate access” need a login.
DefenseLayered infrastructure defense90 seconds: three layers hand over
Air forceRunway repair orderingThe fastest order, proven
RescueRescue LabWildfire, mountain, water, road clearance
Sea rescueMaritime search and rescueFour aircraft find a person overboard
Public safetyPublic Safety LabRehearse the block before entry
Try itResearch accountUrban chaseAI vs AI, or play yourself
BuildBABYLON MarketBuild the vehicle first
DecideSeparate accessZERO staff consoleAlerts become courses of actionStart with these three
No login. About 2 minutes each. Each one starts playing on its own; pause any time. Press “Narration on” for the voice-over.
Street search × fog of war
Satellite imagery of a real Hsinchu block. Same three aircraft, same four minutes, two search methods compared; the fog shows who can see what now and where the picture is stale.
Layered critical-infrastructure defense
Three-layer terminal defense: fused sensing, mobile launch, four moving intercept screens across three layers, reserve re-tasking and layered handoff.
Maritime search and rescue
Four shipborne aircraft search a drifting probability field, independently confirm a person overboard, then release Voyager I to tow a life ring.
Try it yourself
Interactive. Items marked “research account” need a login.
Urban chase: AI vs AI, red vs blue
Compose two teams, watch AIs chase each other or play yourself; every match can be replayed.
Weighted search: intruders, spot fires, people overboard
One search for three missions: probability × value. The targets stay hidden; a conventional pattern flies against the same targets at the same speed, with statistics over 200 placements.
ZERO operations assistant: from alert to courses of action
The defence language-model console, deployable on-prem. It turns an alert and the intelligence into three courses of action and, once approved, an execution plan; numbers come from the ballistics, line-of-sight and viewshed engines, with sources.
Runway repair ordering (ZERO)
How soon can aircraft fly after an attack? The strip and repair order, with the proof that nothing faster exists laid out to check.
Flight physics: the JSBSim sim bus
Three F450s patrol a base on aerodynamics and motor physics at 250 steps a second; change the wind, rerun, and see the same state handed to Isaac Sim, Unreal Engine and DIS.
Fog-of-war console
A pannable, zoomable map with three-state fog, information age and first-person views stitched into a board view.
Reaction forecast (ZERO × ALAYA)
Forecast the other side’s likely reaction, avoid and continue the mission; full replay comparison on the same scene and seed.
Wufeng open terrain
A georeferenced common operating picture and fleet state on real Taiwanese terrain.
Exhibition mode
Mission-preparation screens for wildfire sensing and rescue, built for exhibition displays.
From simulation to real vehicles
Each layer claims only what it has done.
Ten command briefs, fixed seeds, urban-chase replays.
ArduPilot SITL, three Copter 4.7.2 vehicles; one is hit mid-mission by a battery drop, GPS loss, link loss or process kill, 10 runs each, plus 10 fault-free baselines: all 40 faults detected (median 0.4–2.6 s), all 50 missions completed, coverage 0.99. Simulation results; report dated 22 Oct 2026.
Repeat the three-aircraft fault injection and re-tasking on real aircraft running the same ArduPilot flight stack as the SITL runs. The algorithm was first tried indoors on Tello EDU (Wi-Fi-controlled consumer drones).
MAVLink 2 decoding and conformance tests (PX4 / ArduPilot telemetry, receive-only).
Public record: Taiwan-side host of TDTH 2026, the Taiwan Defense Tech Hackathon (2–4 Oct, National Taiwan University); presented Tactical AI for Unmanned Vehicle Teams at Plug and Play Japan, Tokyo (6–7 Oct).
Research and integration records →Equipment and research notes
Rescue Lab
Wildfire, mountains, water, roads: the same engines, rerun in rescue terms.
Public Safety Lab
Pre-entry reconnaissance, missing-person and suspect searches, perimeter watch: the same engines, rerun in police terms.
BABYLON Market
Build the vehicle before you build the pitch.
Research insights
The inference, the limits and the failures in plain sight.
Want to run your own mission?
A research account unlocks urban-chase matches, mission editing and full replays. Demonstrations can be on-site or online.