TATVA · DRONES & UAVS
An on-board AI that works when the link doesn't.
Tatva Edge runs on the drone's companion computer, with no cloud and no network. An operator's plain-language instruction becomes an exact, structured command — and a request outside the mission's scope gets no action at all. Built in India.
In short
Tatva for drones is an on-device AI model from AgentAnywhere that runs on a UAV's companion computer with no internet connection. It turns an operator's plain-language instruction into a structured command and declines requests outside its defined scope. It is built in India and runs fully offline or air-gapped.
- Runs on
- The drone's companion computer or the ground station — offline.
- Does today
- Plain-language instruction → structured command, or no action if out of scope.
- Comes next
- On-device perception. Not claimed today.
- Autopilots
- Flight-stack adapters run so far only against simulators, with commands on ArduCopter 4.7.1 and PX4 v1.18.0-rc1. No real drone has been connected.
Claims last reviewed .
The places drones matter most are the places the cloud isn't.
A survey line over a pipeline corridor, a search grid in a valley after a landslide, a perimeter at a remote site: the link is thin, contested, or absent by policy. A drone whose intelligence lives in someone else's data centre stops being intelligent exactly when it is needed.
Tatva takes the opposite position. The model sits on the drone, the instruction is interpreted on the drone, and nothing — not the imagery, not the command, not the mission — is sent anywhere to be understood. It is the same sovereignty principle as the rest of AgentAnywhere, taken to the edge of the network and then past it.
Five questions a drone has to answer.
Drone AI is five layers, not one. Tatva is honest about which of them it answers today — and the platform around it answers the rest with a human in command.
Perception — what am I seeing?
Detection, tracking, thermal and defect finding. On-device perception is where Tatva Edge goes next; it is not claimed today.
Navigation — where should I go?
Obstacles, terrain, and positioning when GPS is weak. This stays with the flight stack — Tatva does not fly the drone.
Reasoning — what does it mean?
Turning what the sensors report into something an operator can act on, without sending the frame to a cloud to be explained.
Agentic action — what next?
This is what Tatva Edge does today: an instruction becomes one exact, structured command — plan, fly, inspect, revisit, report.
Fleet intelligence — who does what?
Sectors, routes, battery rotation and handover across many drones, with one operator holding the whole picture.
One controller who decides
The drone recommends. A named human approves every consequential action, and a policy outside the model blocks what must never happen.
It plans the action — or it declines.
For a drone, the most important property of an on-board model is not what it can do. It is what it will not do.
In scope
“Inspect solar block C.” One goal in: a coverage plan, the flight, a hotspot found, a request to go lower before it does, and a signed report out.
Every command is one of a defined vocabulary, with named arguments. There is no free-text path from a sentence to a motor.
Out of scope
A request with no matching in-scope action returns nothing — not a best guess, not a creative reading. The model emits valid structured actions or none.
And the model cannot talk its way past the policy. No-fly zones and geofences, keep-out yards, never descending over people: these are checked outside the model, where a persuasive sentence has no effect.
Where drone operators use it.
Infrastructure inspection
Solar farms, power lines, pipelines and towers — sites that are large, remote, and often forbidden to connect.
Survey & mapping
Long corridors and wide areas where the drone is out of link for most of the flight.
Search & rescue
Disaster zones where the network is the first thing to fail and minutes matter.
Perimeter & border
Observe-only patrol inside your own wire, where no frame may leave the perimeter.
Drone OEMs & integrators
A sovereign on-board assistant to ship inside your airframe, under your brand, on your hardware.
Regulated operators
Every recommendation carries a signed receipt you can verify offline — change one character and the receipt fails.
Use cases you can see running.
Each of these runs on our demo hub. The drones and scenes are simulated; every Tatva Edge command shown is real recorded model output. Perception and swarm behaviours are concepts we build with design partners, and each demo says so on screen.
Swarm handover
Five hundred drones over a town. Lose the leader and the swarm elects a new one on board and keeps flying. Send it home and every drone flies its own route to its own landing slot, so losing the leader again costs nothing.
Rooftop solar inspection
A whole neighbourhood of rooftop arrays in one tasked flight. Hot panels are flagged for a named engineer, who approves any closer look.
Flood search and rescue
Drones split a flooded area between them and keep searching when the towers go down. Every finding reaches mission control with a signed record.
Border reconnaissance
Drones patrol a border line and track suspicious movement. A person decides the response: keep tracking, send a ground patrol, or mark a false alarm. Awareness only.
Farm watch
Crop-health maps and pest hotspots across smallholdings, with spraying only where it is needed and only after the farmer or operator approves.
Drone light shows
A thousand drones forming the tricolour, the map of India and a portrait, each moved to its point without crossings, with spares filling any gap. A person starts every formation.
What it runs on.
On the edge box, Tatva Edge runs inside AgentAnywhere Nabhika OS — our sovereign operating-system layer. Nabhika puts a kernel-level network filter and a sandbox around the model runtime, so the model reaches the local broker on the box and nothing else, and it signs a hash-chained record of what is installed. The model cannot leave. The node can prove what it ran.
Nabhika OS is still being built. The edge evidence we show today is a recorded test run on a small single-board-class computer — not yet a drone in flight — and we say so wherever we show it.
See it run.
Drone Mission Ops is our drivable demo suite: a 3D airspace, a fleet board, and three consoles — perimeter patrol, solar-farm inspection, and a three-drone search-and-rescue sweep. You take the controller's seat, approve or reject each recommendation, and verify the receipts yourself.
Alongside it, the demo hub runs Swarm Handover, Border Recon, Farm Watch and Sky Canvas, a thousand-drone light show. In each one you give the command, watch it travel to the drone, approve what matters, and see the signed record come back.
The flights in it are simulated, and every screen says which elements are captured model output, which are real code, and which are synthetic. No drone was flown for the demo. Ask us for a walkthrough.
What Tatva for drones is not.
An honest boundary is part of the product. If a vendor cannot tell you what their drone AI does not do, that is the finding.
- It is not an autopilot. Tatva does not stabilise, steer or land the drone — your flight controller does.
- It is not on-device perception yet. Detection and tracking on the drone are the next step, not a current claim.
- It is not a UAV-specialised model yet. The command vocabulary is defined and the mechanism is proven on a demonstration tool space; the airframe-specific model is being built with design partners.
- It is not a targeting or autonomous-engagement system. It assists a human controller, who decides.
- It has not been connected to a real drone. Our flight-stack adapters, which carry an approved command to a drone's own autopilot, have been run against the ArduPilot ArduCopter 4.7.1 and PX4 Autopilot v1.18.0-rc1 simulators; nothing has been run against a real drone or a real flight controller, and in those runs a script stood in for the software that would call the adapter. Fitting it to your drone is work we do with you.
FAQ
Frequently asked questions.
- Can an AI model run on a drone without an internet connection?
- Yes. Tatva is built to run on-device — on a drone's companion computer or at the ground station — with no network and no cloud. The model and the data stay on the hardware. It is designed for thin, contested or prohibited links, where a cloud-dependent model stops working.
- What does Tatva actually do on a drone today?
- Today, Tatva Edge turns an operator's plain-language instruction into one exact, structured command from a defined vocabulary, and returns no action when a request is out of scope. On-device perception — detection and tracking on the drone — is the next step and is not claimed today.
- Does Tatva fly the drone?
- No. Tatva is not an autopilot. Flight control, stabilisation and navigation stay with your flight controller. Tatva assists the human operator: it interprets instructions and recommends actions, and a named human approves every consequential one.
- How do you stop an on-board AI from doing something unsafe?
- Two ways. The model only emits commands from a defined vocabulary and declines anything outside it. And safety rules — no-fly zones, geofences, keep-out areas, never descending over people — are enforced by a policy check outside the model, so the model cannot argue its way past them. Every recommendation is signed.
- Is Tatva for drones made in India?
- Yes. Tatva is AgentAnywhere's own model family, trained from scratch in India by ShepHertz, and it runs entirely on your hardware inside your jurisdiction. Nothing is sent to a third-party cloud.
- Can we integrate Tatva into our own airframe?
- Fitting it to your drone is work we do with you. Our flight-stack adapters have been run against the ArduPilot ArduCopter 4.7.1 and PX4 Autopilot v1.18.0-rc1 simulators; nothing has been run against a real drone or a real flight controller. Bring your airframe and your mission and talk to us.
Put a sovereign model on your airframe.
We are working with drone OEMs, integrators and operators who need intelligence that survives a lost link. Bring your airframe and your mission.