Rail inspection is one of the few duty cycles where the case for a legged platform is not a matter of preference. The asset that needs inspecting — track bed, tunnel lining, cable troughs, signalling equipment, rolling stock undersides — sits in a space with no GPS, no lighting, restricted clearance, and a hard rule that humans should not be walking there during traffic hours. Fixed sensors cover the points where it is economical to bury cable. A quadruped covers everything else.
This article focuses on what is specific to rail: the environment, the payload, and the sampling frequency the standards actually require.
Why rail is different from a plant patrol
A substation inspection round is a loop: the robot walks the same path, stops at the same 40 gauge points, returns. Rail inspection is not a loop. A tunnel section may be 4 km long; the items to inspect are distributed unevenly; some are 2 m above the track bed and some are at ankle level. The practical consequences:
- Range, not speed, is the binding constraint. A 2-hour endurance robot that covers 1.5 km per charge will not finish a 4 km tunnel with meaningful dwell time at each asset. Either budget for more robots, a mid-route fast-charge point, or a shorter section per night — decide this before the vendor conversation.
- The corridor is narrow and the clearances are fixed. Turning circle, crossing width and body height are hard constraints set by the structure gauge. A platform that is 5 cm too tall or 10 cm too wide is useless regardless of its other merits.
- Exposure is long and repetitive. Track-bed fasteners, cable brackets and tunnel-lining segments repeat thousands of times. The defect classes are visually similar and the detection task is closer to industrial parts inspection than to patrol anomaly detection.
Lidar, thermal, and why you need both
The rail use case splits cleanly into geometry and temperature, and the two sensor families do not substitute for each other.
| Task | Primary sensor | Notes |
|---|---|---|
| Tunnel lining crack / spall / water ingress mapping | 3D Lidar + high-res visible camera | Lidar gives you the geometric deviation; the camera gives you the crack width. Neither alone is defensible in a report. |
| Cable bracket, trough cover, fastener condition | Visible camera, close range | Dwell time and standoff distance matter more than sensor choice; the robot must be able to position consistently within ±5 cm of the same pose every run. |
| Overheating of cable joints, traction power connections | Thermal (LWIR) camera | Requires stable emissivity assumptions and a consistent viewing angle. Absolute temperature readings are only meaningful if the same pose is repeated run to run. |
| Third-rail and OCS clearance verification | Lidar + thermal | Thermal identifies the live conductor; Lidar confirms clearance to the gauge envelope. |
| Standing water, drainage obstruction | Lidar (ground plane) + camera | Cheap to add once you already carry the sensors; a common early win for ROI reporting. |
In practice, the thermal camera is the payload that changes the commercial case. Assets that degrade thermally — joints, connectors, traction returns — are exactly the ones where a missed defect costs the most, and exactly the ones where a manual round in a dark tunnel is least reliable.
Navigation in a signal-dead corridor
Rail tunnels are the canonical GPS-denied environment: long, geometrically repetitive, often without texture, sometimes wet or reflective. Pure Lidar SLAM degrades in exactly this geometry because loop closure has nothing distinctive to close against over long distances. The workable approaches, in the order we prefer them:
- Pre-mapped route with AMCL-style localization against a prior map. Map the tunnel once with a survey-grade rig, then run localization-only during operations. This is the most repeatable option and the one that makes run-to-run comparison of gauge readings and thermal images valid.
- Lidar–IMU fusion with rail-based odometry augmentation. Where the platform can maintain a consistent lateral offset, the rail itself acts as a strong 1-D odometric reference that corrects accumulated longitudinal drift.
- Fiducial markers at defined intervals. Cheap, robust, and unfashionable. In a tunnel you control the infrastructure, so you can install them — this is often the pragmatic answer for a first deployment.
What we do not recommend is running live SLAM as the sole navigation source in a long corridor. The failure mode is silent: pose degrades gradually, the robot keeps walking, and the report comes back with plausible-looking thermal readings taken from the wrong positions.
Hazardous areas and the actual constraints
Not every rail environment is explosive-atmosphere rated, but several are: battery rooms, fuel sidings, and certain maintenance depots fall under hazardous-area classification, and the equipment used there must be certified accordingly. If your deployment crosses into a classified zone, the certification path constrains the platform far more than the robot's performance spec. The rules that matter are covered in what ATEX Zone 1 actually means for a quadruped; the short version is that intrinsic safety and enclosure-based protection both involve mass and heat-dissipation penalties that change the endurance numbers you were quoted.
Scoping a first deployment
A rail pilot that produces decision-grade data, rather than a demo, usually looks like this:
- One section, one shift, one asset class. Pick 800–1200 m of tunnel, or a single depot bay, and inspect one class of asset — lining defects, or cable brackets, not both. The purpose is to establish detection reliability, not coverage.
- Establish the ground truth yourself. Have your own inspectors walk the same section and produce an independent defect list. Without this, the pilot's false-negative rate is unknown, and the false-negative rate is the only number that matters for a safety case.
- Run it at least three times before judging. Repeatability across runs is the actual product. A system that finds 80% of defects once and 55% the next week is worse than one that reliably finds 65% every time.
- Define the data handoff. Decide in advance whether the output lands in your asset-management system, your inspection platform, or a PDF. Integration is where rail pilots usually stall — long after the robot has proven it can walk the tunnel. We cover the interface side in how to write an inspection robot specification.
The economics
Rail inspection is one of the easier cases to justify financially, because the comparison is not robot-versus-human on labour cost — it is inspection-during-traffic-hours versus inspection-in-a-possession. A robot that lets you inspect without taking a possession, or shortens the possession window, is competing against the cost of disrupted service, which is a much larger number. Where the robot patrols during a scheduled engineering window with no service impact, the comparison reverts to labour and the payback period lengthens considerably. Be honest about which of the two you are buying. The full model is in our TCO framework.
What to ask a rail-inspection vendor
- Show recorded pose data from a 30-minute run in a tunnel with your clearance profile — the actual pose estimate, not a video.
- What is the run-to-run positional repeatability at a fixed gauge point? Ask for a number in millimetres.
- How does the platform behave when localization confidence drops? Stop-and-flag, or continue-and-estimate?
- What is the emissivity assumption in your thermal defect thresholds, and how do you handle a wet versus dry tunnel floor?
- What is the manual intervention rate per 100 inspection runs?
An integrator who has actually deployed in tunnels will have all five answers on hand. One who has not will answer the first three and defer the rest.
If you are scoping a rail or transit inspection project, our rail & transit inspection page covers the platform configurations we build for tunnels, depots and track-side assets, and you can send us the section profile for an honest feasibility read.