India has spent a decade building roads at a pace few countries have matched. The harder problem, which arrives quietly once the ribbon is cut, is knowing the condition of what has been built. A road network measured in lakhs of kilometres cannot be inspected on foot, and a pothole that is not recorded is not repaired. AI infrastructure inspection, the use of cameras, sensors and computer vision to survey assets and detect defects automatically, promises to make condition data continuous rather than occasional. For road agencies, contractors and the companies that supply them, the opportunity is real. So is the risk of buying detection software that produces impressive maps and no repaired roads.
The gap is in the workflow. Detecting a crack in an image is a solved research problem to a useful degree. Turning millions of detections into a prioritised, funded, executed and verified repair programme is an institutional problem, and that is where most inspection programmes either create value or stall.
This piece sets out the scale of India's road asset base, the programmes already applying technology to its inspection, what published research says about the accuracy of computer vision for road damage, and an original workflow map covering five stages: survey, detect, prioritise, repair and verify. It is written for officials in road-owning agencies, infrastructure investors and executives in construction and engineering firms.
The asset base: what has to be inspected
Union Road Transport Minister Nitin Gadkari put India's road network at about 63.73 lakh km in June 2023, describing it as the second largest in the world, as reported by Deccan Herald. National highways are a small but critical share. A Ministry of Road Transport and Highways statement reported by IANS in December 2025 put national highway length at 1,46,560 km, up about 61% from 91,287 km in 2014. Highways of four lanes and above rose from 18,371 km to 43,512 km, and access- controlled high-speed corridors and expressways from 93 km to 3,052 km.
Rural roads are the other large category with central oversight. According to Ministry of Rural Development figures reported by NBM and CW in August 2026, the Pradhan Mantri Gram Sadak Yojana has completed 1,85,651 rural roads spanning about 7,96,941 km and 10,833 bridges since its launch. Contractors maintain these roads for five years under performance-based maintenance contracts, maintenance is tracked on the eMARG digital platform, and inspections are carried out by programme units and by state and national quality monitors.
The rest of the network, state highways, district roads and urban streets, sits with state public works departments and municipal bodies, each with its own budget, contractors and record-keeping. That fragmentation matters for inspection: the technology may be common, but the owners, data standards and repair processes are not.
Condition has human consequences. Ministry data compiled by Factly show that accidents attributed to potholes rose to 4,446 in 2022, with 1,856 deaths, though both figures are well below the peak of 11,106 accidents in 2014. Attribution of crashes to road condition depends on police recording and is likely to understate the effect, but the trend shows the problem has not gone away.
What is already being done
The most significant national programme is on highways. News On AIR reported in October 2025 that the National Highways Authority of India would deploy network survey vehicles using a 3D laser- based system across 20,933 km of national highway stretches in 23 states, capturing road inventory and pavement condition data including surface cracking, potholes and patches. The data is uploaded to NHAI's AI-based Data Lake portal, where an expert team analyses it.
This matters for two reasons. First, it establishes that a road owner at national scale now treats machine-collected condition data as the basis for maintenance, rather than periodic manual inspection. Second, it shows the architecture that works: specialised data capture, central storage and expert analysis, connected to the authority that pays for repairs. Network survey vehicles are expensive and cover a defined network. For the far larger network of state, district and urban roads, cheaper capture, such as cameras mounted on ordinary vehicles, smartphones or buses, is the likelier route, with lower data quality and higher frequency.
On rural roads, the combination of a five-year defect liability period and a digital maintenance platform creates a ready structure into which inspection data could feed: if condition can be measured objectively and frequently, performance-based contracts can be enforced on evidence rather than on periodic site visits.
How accurate is computer vision for road damage?
The research base is substantial and, usefully for India, includes Indian roads. The RDD2022 dataset, published in September 2022 by Deeksha Arya, Hiroya Maeda, Sanjay Kumar Ghosh, Durga Toshniwal and Yoshihide Sekimoto, contains 47,420 road images from six countries, Japan, India, the Czech Republic, Norway, the United States and China, annotated for four damage types: longitudinal cracks, transverse cracks, alligator cracks and potholes.
The dataset underpinned the Crowdsensing-based Road Damage Detection Challenge at the IEEE International Conference on Big Data 2022, in which more than 60 teams from 19 countries registered. The challenge summary paper reports that the best solution, an ensemble of YOLO and Faster R-CNN models, achieved an F1 score of 76% on test data combined from all six countries. More recent research continues to push detection, but results on different datasets and metrics are not directly comparable; a 2025 paper proposing an attention-based model reported an average precision of 0.445 on its own diverse road damage dataset, which illustrates how much reported performance depends on the data and metric.
What does an F1 score of around 76% mean operationally? F1 balances precision, the share of detections that are real defects, and recall, the share of real defects that are detected. A system at that level, in images collected from vehicles in mixed conditions, will both miss some defects and flag some non-defects. For a survey of thousands of kilometres, that produces very large numbers of both. The engineering response is not to wait for a perfect model but to design the workflow so that each type of error is caught at the stage where it costs least: false positives filtered before repair crews are dispatched, missed defects caught by repeat surveys and citizen reports.
Three further limits matter. Image-based detection sees the surface, not the structural condition beneath it, which laser profiling and other sensors address. Performance varies with lighting, rain, shadows, dust, water-filled potholes and the many surface types on Indian roads. And a defect's severity, which drives priority, is harder to estimate from a single image than its presence.
A workflow map for AI infrastructure inspection: survey, detect, prioritise, repair, verify
The original contribution of this piece is a workflow map that sets out, for each of five stages, what AI does, what people do, the characteristic failure, and the metric that shows whether the stage is working. The premise is that value is created only when all five stages connect.
| Stage | Inputs | AI role | Human role | Characteristic failure | Metric that matters |
| Survey | Vehicle-mounted cameras, laser profilers, smartphones, drones for bridges and structures | Quality checks on captured data; flag blurred, dark or incomplete runs | Plan routes and frequency; ensure coverage of the owner's whole network | Patchy coverage; data collected but not geotagged to the asset register | Share of network surveyed in the period, at usable quality |
| Detect | Images, point clouds, location data | Detect and classify defects; estimate extent and severity | Review a sample; validate new defect classes and new road types | Model trained on other roads underperforms on local surfaces and conditions | Precision and recall on locally labelled samples, by defect type |
| Prioritise | Detections, traffic volume, road class, accident history, budget | Rank defects and road sections by risk and cost; group into work packages | Set the policy weights; approve the programme; handle exceptions | Ranking ignores safety or traffic, or budget is set before the data | Share of high- risk defects in the funded programme |
| Repair | Work orders, contractor assignments | Generate work orders with location, photographs and quantities | Contractors execute; engineers supervise | Detections never become work orders because systems are not connected | Time from detection to work order; time from work order to repair |
| Verify | Post-repair survey, contractor reports | Compare before and after imagery; detect recurrence | Accept or reject work; enforce contract terms | Repairs marked complete without evidence; defects return within months | Share of repairs verified by imagery; recurrence rate within the defect liability period |
Reading across the stages explains why many technology pilots disappoint. A vendor demonstration usually covers survey and detect. The value lies in prioritise, repair and verify, which depend on the road owner's processes, contracts and budget cycle, and which the vendor does not control.
The verify stage deserves particular emphasis in India. Performance-based maintenance contracts, such as those used in PMGSY with a five-year defect liability period, depend on proving whether a contractor met the standard. Repeated, time-stamped imagery of the same stretch gives the road owner objective evidence of condition before and after repair and of recurrence within the liability period. That converts inspection technology from a cost into a contract enforcement tool, which changes who will pay for it.
Beyond roads
The same workflow applies to other linear and structural assets: bridges, culverts, railway track and overhead equipment, canals, pipelines and power transmission lines. Detection inputs differ, drones and close-range imagery for bridge decks and piers, track-mounted cameras for rail, and the consequences of a missed defect are often more severe. For structures where a failure is catastrophic, the balance shifts towards recall over precision: it is better to send an engineer to check a hundred false alarms than to miss one real crack. The workflow map's emphasis on locally measured accuracy by defect type, and on human acceptance of repairs, applies with greater force.
Vision-language models, which can describe what an image shows in text and answer questions about it, add a new capability to the detect and prioritise stages. They can produce a written description of a defect for a work order or answer an engineer's question about a stretch of road. PanScience reports that its venture RoadVision AI applies vision-language models to road infrastructure (company- reported). The usefulness of such systems, like that of narrower detectors, depends on the workflow around them: descriptions that feed work orders and verified repairs create value; descriptions that sit in a dashboard do not.
What road owners and contractors should ask
For a road-owning agency procuring an inspection system, the workflow map translates into questions:
1. Coverage: what share of our network will be surveyed, how often, and at what data quality? 2. Local accuracy: what are precision and recall by defect type on a labelled sample of our own roads, not on a public benchmark? 3. Integration: how do detections reach our asset register and work order system, and who owns the data? 4. Prioritisation: can we set our own weights for safety, traffic and cost, and see why a section was ranked as it was? 5. Verification: will the system re-survey repaired sections and report recurrence within the defect liability period? 6. Data ownership and standards: can we take the data to another vendor, and does it follow a common schema across our divisions?
For contractors, objective condition data is double-edged. It exposes poor work, but it also protects contractors who meet the standard from disputed assessments. Contractors who adopt inspection technology for their own quality control can enter acceptance discussions with evidence of their own.
The financing logic
Inspection programmes are often justified as a safety measure and funded as a technology project, which makes them vulnerable when budgets tighten. A more durable logic links them to maintenance spending. If objective data helps an agency direct a fixed maintenance budget to the defects that matter most, and enforce contractor obligations on roads already under defect liability, the inspection cost is recovered from avoided waste rather than added to the budget. The data from NHAI's network survey programme, from PMGSY's eMARG platform and from municipal pilots would allow such a case to be tested; the important design choice is to measure, from the start, the metrics in the right-hand column of the workflow map.
A separate PanScience Research piece covers evaluation before deployment, and a separate Cluster 4 piece on AI observability covers how to monitor models once they are live. Both apply: a detection model that performs well in the first year can degrade as cameras, seasons and road surfaces change.
Conclusion
AI infrastructure inspection in India has moved from research to national practice, led by NHAI's survey programme on highways and supported by a research base that includes Indian roads. Detection accuracy is good enough to be useful and not good enough to be trusted without a workflow around it. The value of an inspection programme lies in connecting survey and detection to prioritisation, repair and, above all, verification, so that the data changes what is fixed and holds contractors to the standards they have agreed. Agencies that buy detection alone will get better maps. Agencies that design the whole workflow will get better roads.
Methodology and sources
This article combines government data, press reports of official programmes, peer-reviewed and preprint research, and an original workflow map. Road network length is from Deccan Herald's report of Minister Nitin Gadkari's statement, 27 June 2023 (deccanherald.com, story 1231686). National highway figures are from IANS, reporting a MoRTH statement, 30 December 2025 (ianslive.in/indias- highway-network-crosses-146-lakh-km-to-become-2nd-largest-in-world--20251230180903). PMGSY figures are from NBM and CW, 3 August 2026, citing the Ministry of Rural Development (nbmcw.com/news/roads-highways/pmgsy-completes-nearly-7-97-lakh-km-of-rural-roads-and-10-833- bridges-across-india.html). Pothole accident data are from Factly's compilation of MoRTH data (factly.in/data-fatality-rate-in-accidents-on-potholed-roads-increased-from-0-27-in-2013-to-0-42-in- 2022). The NHAI network survey vehicle programme is from News On AIR, 22 October 2025 (newsonair.gov.in/nhai-to-deploy-3d-survey-vehicles-in-23-states-for-ai-based-highway-monitoring). Research results are from Arya et al., RDD2022 (arxiv.org/abs/2209.08538), Arya et al., CRDDC 2022 summary (arxiv.org/abs/2211.11362), and Alkalbani et al., 2025 (arxiv.org/abs/2501.02822). The workflow map, the interpretation of F1 scores for operations and the financing logic are the author's analysis; no inspection deployment is described as observed. The RoadVision AI description is company-reported by PanScience Innovations (panscience.xyz).