How Mobile Robots Are Used in Bridge and Infrastructure Inspection

TL;DR: Robot infrastructure inspection can replace some rope-access surveys and scaffolding with mobile ground robots, drones, and crawlers that carry cameras, LiDAR, thermal sensors, and ultrasonic probes. On bridges, pipelines, tunnels, and dams, unmanned ground vehicles (UGVs) like the Leo Rover let engineers capture geo-tagged imagery, point clouds, and corrosion data with fewer traffic closures and less need to send technicians into confined spaces. ROS 2 makes sensor fusion, autonomy, and data logging reproducible across deployments. This article explains how the systems work, what sensors they need, and where they have been used in real projects.

Why Robot Infrastructure Inspection Matters

Civil infrastructure ages faster than inspection budgets grow. The 2021 ASCE Report Card graded U.S. bridges at C and noted that 42% of all bridges are at least 50 years old, with 7.5% classified as structurally deficient. Visual inspection by certified engineers remains the legal baseline in most countries, but it is slow, expensive, and exposes people to traffic, heights, and confined spaces.

Robot infrastructure inspection addresses three concrete problems. First, access: many structural elements (box girders, pier caps, pipeline interiors, dam galleries) are physically hard to reach. Second, repeatability: human inspectors produce subjective reports that are difficult to compare across years. Third, downtime: closing a highway bridge for a routine survey can cost operators tens of thousands of euros per day. Mobile robots can reduce all three.

The technology is no longer purely experimental. Network Rail in the UK and several U.S. state DOTs have used UAVs in inspection workflows, while agencies such as the Swiss Federal Roads Office have evaluated remote-sensing and digital inspection methods. UGVs and crawlers are used more selectively, especially where close-range sensing or confined-space access is needed. Research groups at ETH Zürich, Carnegie Mellon, and TU Delft have published field deployments on bridges, tunnels, and offshore structures.

What Types of Robots Are Used in Infrastructure Inspection

No single platform covers every structure. Inspection robotics splits roughly into four categories, each suited to a different access problem.

• Unmanned ground vehicles (UGVs): wheeled or tracked robots that traverse decks, service walkways, tunnels, and pipeline corridors. Useful when payload, runtime, and stable sensor mounting matter more than vertical reach.
• Unmanned aerial vehicles (UAVs): drones for the underside of bridges, towers, chimneys, and roofs. Limited by wind, GPS denial under decks, and short battery life (typically 20-35 minutes).
• Climbing and magnetic crawlers: robots that adhere to steel surfaces using magnets or vacuum, used on storage tanks, ship hulls, and steel box girders.
• In-pipe robots: tethered or autonomous crawlers for water mains, sewers, and gas pipelines. Often equipped with pan-tilt cameras and sonar.

UGVs occupy a useful middle ground for field robotics: more payload than a drone, more sensor stability than a crawler, and easier to operate than a manipulator-equipped platform. The Leo Rover, for example, carries up to 5 kg of payload, supports ROS 2 integrations, and exposes a standard mounting plate and power rail for adding LiDAR, thermal cameras, or GNSS-RTK receivers.

How a Typical Bridge Inspection Mission Works

A robot infrastructure inspection mission usually follows a repeatable workflow. The point of automation is not to remove the engineer, but to give them better data to interpret.

1. Pre-mission planning: the inspection target is segmented into zones (deck, piers, abutments, bearings). Waypoints or coverage paths are generated, often in QGIS or a ROS 2-based planner.
2. Localization setup: outdoors, GNSS-RTK provides centimeter-level positioning. Under a bridge deck or inside a tunnel, the robot switches to LiDAR-inertial odometry (for example, FAST-LIO2 or LIO-SAM).
3. Data capture: the robot follows the path while logging synchronized streams (RGB, thermal, LiDAR, IMU, GNSS) into a ROS 2 bag.
4. Post-processing: point clouds are registered into a 3D model; images are stitched and run through crack-detection networks (U-Net variants are common in published work).
5. Reporting: defects are geo-tagged, severity-classified, and exported to the asset management system (e.g., Bentley AssetWise, IBM Maximo).

What Sensors a Mobile Robot Needs for Infrastructure Inspection

Sensor selection is driven by the defect catalog the inspection has to detect. For concrete bridges, the relevant defects are cracks (>0.2 mm often reportable), spalling, delamination, and corrosion of exposed rebar. For steel, the focus shifts to coating failure, pitting, weld defects, and section loss.

• High-resolution RGB camera: at least 12 MP, global shutter preferred. To resolve a 0.2 mm crack reliably, you typically need a ground sampling distance of about 0.07-0.1 mm per pixel, which sets the standoff distance for a given lens.
• Thermal camera: uncooled microbolometers (for example FLIR Boson-class modules; lower-resolution Lepton modules can be used for screening) detect delamination in concrete via solar-loading thermography. Resolution of 320×256 or higher is typical for defect mapping.
• 3D LiDAR: Livox Mid-360, Ouster OS0, or Velodyne VLP-16 class sensors for geometry capture and SLAM. Range and point density matter more than channel count for inspection.
• GNSS-RTK: u-blox ZED-F9P or equivalent, paired with an NTRIP correction service for georeferencing.
• IMU: industrial- or tactical-grade preferred for LiDAR-inertial fusion under decks where GNSS drops out.
• Optional: ultrasonic or GPR probes for subsurface concrete assessment, mounted on a deployable arm.

ROS 2 (Humble or Jazzy) is a common integration layer. Each sensor publishes on a standard topic, the tf2 tree describes mounting offsets, and rosbag2 records everything with hardware timestamps. This is the same kind of stack used in academic research deployments, which makes results easier to reproduce and benchmark.

Real Examples: Bridges, Pipelines, and Tunnels

Several documented projects show what works in the field and what does not.

Bridges. The SPRINT project at the University of Nevada, Reno, used a wheeled robot with GPR and impact-echo sensors to map deck delamination on in-service bridges, reporting survey speeds roughly 5-10 times faster than manual chain-drag methods (La et al., Automation in Construction, 2013). Swiss firms and public agencies have also demonstrated drone-based photogrammetry on viaducts where the resulting 3D models are compared year over year to detect geometric change.

Pipelines. Pure Technologies’ PipeDiver and SmartBall tools have inspected thousands of kilometers of large-diameter water mains using inline acoustic sensors. For external above-ground pipeline corridors, UGVs with thermal and gas-leak sensors are used to walk service roads in petrochemical sites.

Tunnels. ETH Zürich’s Robotic Systems Lab has deployed legged and wheeled robots in underground and tunnel-like environments for mapping and defect localization, publishing results that show LiDAR-inertial SLAM achieving low drift over hundreds of meters in GNSS-denied environments.

Dams and locks. The U.S. Army Corps of Engineers has trialed underwater and surface robots for lock-gate inspection, reducing the need for dewatering, which is one of the most expensive operations in dam maintenance.

What these projects share is a focus on reproducible data capture. The robot is a sensor-positioning system; the value sits in the dataset and the analytics that turn it into a maintenance decision.

What Platform Requirements Matter Most in Field Robotics

Choosing a base platform for inspection work comes down to a small number of engineering constraints. The platform has to survive the environment, carry the sensors, and integrate with the software stack the team already uses.

• Ingress protection: at least IP54 for outdoor bridge work; IP65 or better for wet tunnels and coastal sites.
• Payload and power: 3-5 kg payload and a 12 V / 5 V regulated rail cover many LiDAR plus thermal plus compute configurations.
• Compute: an onboard SBC capable of running ROS 2 with at least one sensor SLAM node in real time. NVIDIA Jetson Orin Nano or Xavier NX class is typical.
• Open software stack: ROS 2 support with documented APIs avoids vendor lock-in and lets your team modify behaviors.
• Field serviceability: swappable batteries, accessible wiring, and CAD files for custom mounts reduce downtime.

Open-source platforms like the Leo Rover are designed around these constraints, which is why they appear in university labs and small inspection service companies. You can see more application notes on the inspection applications page.

FAQ

Can a mobile robot fully replace a certified bridge inspector?

No. In nearly every jurisdiction, the legal inspection report must be signed by a qualified engineer. Robots collect data; engineers interpret it. The realistic gain is faster data capture and better year-over-year comparison, not the removal of human judgment.

How accurate is robot-based crack detection?

Published deep-learning crack detectors on concrete (typically U-Net or DeepLab variants) report F1 scores in the 0.85-0.95 range on curated datasets. Field performance is lower because of lighting, wet surfaces, and biological growth. Treat automated detection as a triage layer, not a final classification.

What is the typical operating time for an inspection UGV?

Most small UGVs in this class run 2-4 hours per battery depending on payload and terrain. Swappable batteries are standard. For a full-day bridge survey, plan for two to three swaps.

Do you need RTK GPS or is standard GNSS enough?

For defect geolocation that has to be revisited and compared between surveys, RTK is effectively required. Standard GNSS gives 2-5 m horizontal error, which is too coarse to relocate a specific crack on a pier.

How does the robot localize under a bridge deck where GPS does not work?

LiDAR-inertial odometry is the standard answer. Algorithms like FAST-LIO2 or LIO-SAM fuse LiDAR scans with IMU data, while LiDAR odometry methods such as KISS-ICP can also support localization in structured environments. Drift is often kept under 1% of trajectory length when the scene has enough geometric structure.

Is ROS 2 mature enough for production inspection work?

Yes. ROS 2 Humble (LTS until 2027) and Jazzy (LTS until 2029) are used in commercial deployments. The DDS middleware supports multi-sensor data flow with QoS controls, and tools like rosbag2 give you reproducible logs that hold up to engineering review.

What does it cost to build an inspection-capable UGV?

An open-source base platform plus LiDAR, thermal camera, RTK GNSS, and a Jetson-class computer typically lands in the EUR 15,000-30,000 range depending on sensor choices. Commercial turnkey inspection robots cost several times more.

Where to Go Next

If you are scoping a robot infrastructure inspection project, the next step is matching the defect catalog to the sensor payload and the environment to the platform IP rating. Fictionlab’s team documents reference configurations and field results on the inspection applications page, including sensor integrations and ROS 2 packages you can reuse.