Quadruped and humanoid robots that patrol, inspect and report
Quadruped and humanoid robots for security, industrial plants, commercial buildings and infrastructure. Operators give orders in plain language, the robot localizes without GPS on a 3D map of the site, and every round comes back with photos, positions and events. The AI runs on your own server, inside your perimeter.
Autonomous perimeter round along a fence line
A quadruped robot follows a fence line at dusk and checks it sector by sector.
Robots for every kind of site
Security & surveillance
Perimeter rounds, gates, night patrols and alerts to the control room.
See the use cases →
Industrial plants & warehouses
Inspection rounds of gauges, panels, exits and extinguishers, with photo evidence.
See the use cases →
Retail & commercial buildings
Malls, offices and hotels: after-hours rounds, spoken notices and checks in public areas.
See the use cases →
Infrastructure & utilities
Substations, solar fields, ports and tunnels: wide, unmanned sites with no GPS indoors.
See the use cases →
How do you give orders to a robot dog?
In plain language, in a chat on the dashboard: “do the round of the north fence”, “go to the main entrance”, “turn until you see a person”. A language and vision model on your server splits the order into robot actions, a safety layer checks each one, and the robot reports the outcome.
- Rounds and patrols. To a point, along waypoints or in a repeated patrol.
- Relative moves. “Move forward 3 m”, “turn 90°”, “get closer to the door”.
- Questions. “Where did you see the extinguisher?” “What did you do an hour ago?”
Indoor halls by day, perimeter fence by night
One robot covers a busy hall and, after closing, the fence line outside.
Under the hood
Robot fleet management from one web dashboard
The same software runs quadruped and humanoid robots, with one configuration file per robot. A language and vision model turns the operator’s order into robot actions; a rule-based safety layer checks each action before it starts.
- Orders in plain language. An operator writes “do the round of area B” or “go to the main gate” in chat, in English or Italian. The system splits it into robot actions and reports the outcome.
- One web dashboard. Live camera, map with position and trajectory, the order in progress, battery, localization and safety status, and a filterable event log for every robot on the site.
- Positioning without GPS. The robot localizes on the 3D map of the site with OVER’s visual positioning (VPS). Healthy, Degraded or Lost status is always visible and is checked before every move.
- Rounds and patrols. To a point, along a sequence of waypoints or in a repeated patrol, avoiding obstacles. No-go zones drawn on the map are respected.
- Open-vocabulary detection. The robot finds what you name, such as people, extinguishers, ladders, open doors or equipment, and can keep watch while it moves: on a match it stops and alerts.
- Evidence and site memory. Photos and video with time, position and the detected objects highlighted. The system remembers places, sightings and tasks: “where did you see the extinguisher?”
How does an autonomous patrol robot find its way without GPS?
The site is captured once with a 360° camera into a metric 3D map. The robot then localizes from its own camera images with OVER’s visual positioning system and keeps relocalizing as it moves. Its status, Healthy, Degraded or Lost, is visible on the dashboard and checked before every move.
- Validated against LiDAR. 3.9 cm median surface agreement; metric scale within 0.81% mean error.
- Indoors and outdoors. Basements, warehouses, tunnels and yards, where GPS does not reach.
- No hardware to install. No beacons, tags or markers along the routes.
A 3D map of the site built from the robot’s own camera
A Unitree Go2 walks an office corridor; its camera images rebuild the site in 3D for positioning.
Safety
The AI proposes, the rules decide
The model never moves the robot directly. Every action passes checks that do not depend on the model, and people keep the last word.
- A deterministic check before every action. Localization, navigation state, emergency stop, battery and allowed map points are checked by rules, not by the AI model. Each action is allowed, refused or sent to a person for approval.
- Automatic reflexes. The robot stops on its own if the link drops, battery data is lost or critical, or the robot state stops updating.
- Emergency stop. Always available on the dashboard and independent from the model. Only a named person can release it.
- “Stop” in chat. A stop order written while the robot moves is recognized in about 0.1 s and cancels the motion.
Where does the AI run, and who can switch it off?
On a GPU server inside your perimeter. Language and perception models run on site, with no mandatory cloud service, so camera images and maps stay on your premises and the system keeps working isolated from the internet. OVER is an Italian company under EU law, and only you can switch the system off.
| Question | OVER robots | Typical cloud AI service |
|---|---|---|
| Where does the AI run? | Inside your perimeter | In the provider’s data centre, often abroad |
| Whose models are they? | OVER’s own, installed on site | A third-party lab’s |
| Where do images and data end up? | On your hardware, on your premises | In the provider’s cloud |
| Does it work if the connection drops? | Yes, even fully isolated | No, it needs the cloud |
| Who can switch it off? | Only you | The provider, remotely |
What does a pilot prove on your site?
A pilot runs your own rounds on your own site for about 90 days, so results are measured where the robots will work. You see the robots patrol, inspect and report on the real routes, with the dashboard in your control room and the AI on your server, before extending to other sites.
- Your routes. The robots run the rounds you choose, on the real ground, doors and halls of your site.
- Your control room. Operators give orders, follow every robot and receive alerts on the dashboard from day one.
- Your data. Models, video and maps run on the server installed on your site.
- Measured results. Coverage, rounds completed and events documented, measured on your site before you extend.
Which robots does the platform support?
Today a Unitree Go2 quadruped, for long rounds on rough ground, and a Booster T1 humanoid, for halls and entrances at standing height. Both connect through ROS 2, and the same software drives each one through its own configuration file. A simulator tests missions and regressions without the robot.
How to start
One pilot site, then all the others
- 01
Site survey and 3D capture
We walk the site with a 360° camera and build the 3D map the robot localizes on. We check network coverage (Wi-Fi or 4G/5G), ground and lighting.
- 02
One pilot site
Robots, dashboard and the on-site server run your rounds for about 90 days, with results measured on your site.
- 03
Roll out to more sites
Add robots and sites with the same software and the same dashboard, without redesigning anything.
Questions buyers ask about robots on site
Which languages can operators use?
Orders are written in plain English or Italian in the dashboard chat. Spoken notices from the robot use text-to-speech.
Do we need beacons, markers or GPS on site?
No. The robot localizes with its own camera on a 3D map of the site, built once with a 360° camera walk. It needs network coverage, Wi-Fi or 4G/5G, to reach the on-site server.
How is footage of people handled?
Photos and video are kept for a period you configure, then deleted. Face recognition is optional, limited to registered people, uses an encrypted gallery and is enabled only where the law allows it. Data stays on your premises.
Can one operator follow several robots?
Yes. The dashboard shows every robot on the site on one page, with live camera, position, battery and events, and orders go to each robot from the same chat. Every order, call and event is logged.
We build our own robots. Can you train them on our site?
Yes. The same 3D capture becomes a sim-ready digital twin for VLN/VLA model and policy training. See custom data collection for robotics.
How does a pilot start?
With a site survey and a 360° capture of the area, then about 90 days on one pilot site with your own rounds and results measured there. Then the same setup extends to the other sites.
Project scoping
Plan a robot pilot on your site
Tell us where the robots would work and what they should do. A few clicks, then your contact details. Rough answers are fine.
What happens next
- 01Answer a few questions. Two short steps. Rough answers are fine.
- 02Our reply. We reply by email within 1 business day to set up a short scoping call.
What happens next
- 01Answer a few questions. Two short steps. Rough answers are fine.
- 02Our reply. We reply by email within 1 business day to set up a short scoping call.