Industrial inspection robots for plants and warehouses
Industrial inspection robots for plants and warehouses: a quadruped robot walks your inspection round, photographs gauges, panels and exits, checks that extinguishers and emergency routes are in place, and logs every finding with time and position. Industrial data stays on your own server.
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.
What can industrial inspection robots check on a round?
On a set round, industrial inspection robots stop at each point, take photos and describe the scene: gauges, panels, valves, exits, extinguishers and blocked aisles. Open-vocabulary detection finds the objects you name, and each finding is stored with time and position, so you can compare one round with the last.
- Inspection rounds. A sequence of points across lines, utility rooms and yards, repeated each shift.
- Safety checks. Extinguishers, exits and emergency routes present and not blocked.
- Out-of-hours rounds. Plants and warehouses checked at night and at weekends, without staff on site.
- Ask the site memory. “Where did you see the ladder?” or “what did you do an hour ago?”
Gravel, slopes and rocks, with no operator
Legged robots keep going where wheels stop: loose ground, steps and slopes.
Why a quadruped robot for inspection instead of a wheeled one?
Quadruped robot inspection reaches what wheels cannot: steps, gratings, ramps, cables on the floor and outdoor yards between buildings. The robot localizes with its camera on a 3D map of the plant, so metal structures and indoor areas without GPS are not a problem. Humanoids add reach at standing height.
- A 3D map of the plant. Built from a 360° camera walk; validated against LiDAR at a median of 3.9 cm.
- No-go zones. Areas near machines or forklifts drawn on the map and kept out of every route.
- Change over time. The same round, the same points: what moved, what appeared, what is missing.
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.
Under the hood
One platform for every robot on the site
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?”
Is an autonomous inspection robot safe around people and machines?
Every action passes a deterministic safety layer first: localization, navigation state, battery, emergency stop and allowed map points. Actions are allowed, refused or sent to a person. The robot stops on its own if the link drops, and the dashboard emergency stop works independently from the AI model.
- Human approval. Actions the rules flag go to an operator before the robot moves.
- Stop from chat. A “stop” order cancels the motion in about 0.1 s.
- Release by name. Only a named person can release the emergency stop.
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.
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.
Industrial inspection robots: questions plant managers ask
Can a robot dog for industrial inspection read gauges?
It photographs each gauge or panel from the same point on every round and describes the scene, so values and changes are documented and comparable round after round.
Does it need beacons, markers or Wi-Fi positioning?
No positioning hardware. The robot localizes with its own camera on the 3D map of the plant. It needs network coverage, Wi-Fi or 4G/5G, to reach the on-site server and the dashboard.
Can we train our own robots on our plant?
Yes. The same 3D capture becomes a sim-ready digital twin for training VLN/VLA models and robot policies. See custom data collection for robotics on /robotics-data.
Does the robot need a cloud service?
No. Language and perception models run on a GPU server on your site, with no mandatory cloud service, so video and data stay on your premises. Robots connect to it over the site network through ROS 2.
Which robots does the platform run on?
Today a Unitree Go2 quadruped and a Booster T1 humanoid. The same software drives different robots through one configuration file each, and a simulator lets us test missions without the robot.
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. Tell us the site, the area and what the robot should do, and we scope it with you.
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.
The same robots, in other sites
Security & surveillance
Perimeter rounds, gates, night patrols and alerts to the control room.
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 →