Simulated · CGI
AI FOR MACHINES · DRONES AND UAVS
A drone that still understands you when the link is gone.
Our models run on the drone and the ground station, offline. An operator speaks a plain instruction; it becomes one exact command, a policy outside the model checks it, and a person approves before anything flies.
IN SHORT
AgentAnywhere's AI for drones runs on the drone's companion computer and the ground station with no cloud connection. Tatva Edge turns an operator's plain-language instruction into one structured command or declines it, Sanjaya watches fleet health, and Drashta reads imagery for an analyst. A policy outside the model checks every command and a named person approves anything consequential.
- Runs offline, on the machine
- Sized for constrained hardware
- Approved by the pilot in command
THE PROBLEM
The places drones matter most are the places the cloud isn't.
A survey line over a pipeline, a search grid after a flood, a perimeter at a remote site: the link is thin, contested, or not allowed. A drone whose intelligence lives in a distant data centre stops being useful exactly when it is needed.
So the model has to be on the drone, understand the operator in their own words, and refuse what is out of scope. And a person still has to be in command of anything that matters.
ON THE DEVICE · OFFLINE
What runs where.
Three families, each with one job, all on hardware you own. None of them needs a network to do it.
Language to one machine command — or none.
Turns a spoken or typed instruction into exactly one flight command — take off, survey this block, return home — or no action if it is out of scope.
Where · Tatva Edge on the drone's companion computer or the ground station, offline.
Fleet and telemetry intelligence — the warning before the failure.
Designed to watch battery, link and subsystem health across the fleet and warn before a fault grounds a drone.
Where · On the ground station, next to the fleet board. No telemetry leaves the site.
Machine vision — finds aircraft and buildings in satellite imagery; an analyst decides.
Finds aircraft and buildings in satellite imagery and lists them for an analyst to review.
Where · On a ground workstation, air-gapped. The analyst decides what a finding means.
DEMOS · LIVE ON THE DEMO HUB
Drone demos you can open today.
Every one of these runs on our demo hub. The drones and scenes are simulated; the Tatva Edge commands in them are real recorded model output, and each demo says on screen which parts are which.
Simulated · CGI
Swarm Handover
A swarm of hundreds of drones flies over a town in formation. Fail the leader and the swarm elects a new one on board and re-forms. Speak a command from mission control: it becomes one exact command, policy checks it, a person approves.
Leader handover is a concept; we are looking for design partners for it.
Simulated · CGI
Drone Mission Ops
One operator, one copilot, a whole airspace in 3D. Drive three missions — a solar-farm inspection, a search-and-rescue sweep and a perimeter patrol — approving or rejecting each recommendation, then verify the signed receipts yourself.
Flights are simulated; the policy, referee and receipt code is real.
Simulated · CGI
Border Recon
Three drones keep watch on a fence line from before dawn to sunrise. When something moves, a drone keeps its camera on it and mission control gets an alert with three choices for a person. Awareness and monitoring only.
Detection and tracking are a perception concept; we are looking for design partners for it.
Simulated · CGI
Farm Watch
Survey drones map crop stress over small plots between a canal and a village. The planner proposes spraying only around the pests and keeps clear of the water; the farmer approves before the sprayer lifts off.
Crop-health detection is a perception concept; we are looking for design partners for it.
Simulated · CGI
Sky Canvas
A drone light show paints the Tiranga, the map of India and a portrait of Mahatma Gandhi over a riverfront. Every move is planned and checked before anything flies, and nothing moves until a person presses the button.
Illustration
Sense, reason, act
A drone sweeps a terrain scene as an orbitable 3D theatre. Drashta's findings appear as the sensor passes, Sanjaya raises an early warning on the drone's health, and the command model runs alongside.
The 3D scene is a visualisation; the detections, health readings and commands are recorded model output. The health signal is a replay of a public benchmark, mapped onto this drone's health to show the idea.
Illustration
Tatva Micro in your browser
Not a replay: the on-device command model loads into your browser tab and runs there. Type a drone command and it compiles one exact action, or declines. Turn off Wi-Fi and it still works.
The demo hub is behind sign-in. If you don't have access yet, ask us — we'll set you up, or walk you through it live.
Request demo accessFROM THE APPROVED COMMAND TO THE AUTOPILOT
The last step: carrying the command to the drone's own autopilot.
Our flight-stack adapters carry a command to a drone's own autopilot, which does the flying. All eight of their commands were accepted by the ArduPilot ArduCopter 4.7.1 simulator and by the PX4 Autopilot v1.18.0-rc1 simulator, a release candidate, from take-off to landing. Nothing has been run against a real drone or a real flight controller.
A command goes out only with an approval
The adapter sends a command only if it carries an approval with two valid signatures: one made with the policy check's key and one made with the key of a named person. Without one it refuses, names its reason, and sends nothing to the drone. In the runs so far the signatures were made with test keys.
The lost-link behaviour, in plain words
It reads what the drone is set to do if it loses the ground-station link, the radio-control link or its position, and says it in a plain sentence beside the raw settings. It writes nothing.
A signed record
Every command offered to the adapter, every refusal and every answer from the autopilot is written to a record. Each line is chained to the line before it and signed, so a change to any line is found when the record is checked. The record proves what it contains; it does not prove that nothing followed. In the runs so far it was signed with a test key.
One run on each simulator, on one simulated multirotor. A script stood in for the software that would call the adapter, so the chain from an operator's words to the autopilot has not been run as a whole. The adapter has no arm command; arming stays with the operator. ArduPilot and PX4 Autopilot are independent projects, and ShepHertz is not affiliated with, sponsored by or endorsed by either. ArduPilot is a registered trademark in the United States. The PX4 trademarks are held by Dronecode. The other project names belong to their respective owners. Fitting the adapter to your drone is work we do with you.
Read what a simulator and the source showed usHOW EVERY MACHINE STAYS IN HUMAN COMMAND
The model proposes. It never decides.
On this machine, the named person is the pilot in command.
Policy outside the model
Rules are written and versioned outside the model and checked on every proposal. A model cannot talk its way past them.
A named person approves
Anything consequential waits for a named person. Their approval, or their refusal, is part of the record.
Signed receipts
Every proposal, check and approval is written to a signed, hash-chained receipt that can be verified offline, later, by someone else.
WHAT WE DO NOT CLAIM
The boundary, stated.
FAQ
Frequently asked questions.
- Can an AI model run on a drone without an internet connection?
- Yes. Tatva Edge is sized for the kind of companion computer a drone already carries and runs with no network at all. It turns an operator's plain-language instruction into one structured command or declines it. Nothing is sent to a cloud to be understood.
- Who approves what the drone does?
- A named person. The model only proposes a command; a policy written outside the model checks it against the mission's rules and the pilot in command approves anything consequential. Every step is written to a signed receipt.
- Are the drone demos real flights?
- No. The drones and scenes are simulated and every demo is labelled as such. The Tatva Edge commands inside them are real recorded model output, and the policy, referee and receipt code runs for real on the scenario.
- Has this been connected to a real drone's autopilot?
- No. Our flight-stack adapters carry an approved command to a drone's own autopilot, and all eight of their commands were accepted by the ArduPilot ArduCopter 4.7.1 simulator and the PX4 Autopilot v1.18.0-rc1 simulator, a release candidate. Nothing has been run against a real drone or a real flight controller. The adapter has no arm command; arming stays with the operator.
- Is this a weapons or targeting system?
- No. Tatva and Drashta are not targeting or engagement systems and are never placed in a weapons-release chain. The drone use cases are awareness, inspection, mapping, search and rescue, agriculture and light shows.
- How do I see the drone demos?
- The demo hub is behind sign-in. Request demo access from this page and we will set you up, or book a walkthrough on your own mission.
Claims last reviewed . Demos are simulated demonstrations; each one says on screen which parts are real model output.
Bring us your drones and UAVs.
Tell us the machine, the mission and the hardware. We'll show you the demo on your scenario, and tell you honestly what is ready and what we would build with you.