Published on 26-Aug-2026
Will a Robot Ever Understand the Feel of an Inspection?
Table of Contents
- Session Overview
- Meet the Panel
- Can a Robot Replicate the Human Touch?
- The Sensor Problem: Built for Hands, Not for Robots
- Using the Process Sensor for Its Own Positioning
- The Data Flood and What to Do With It
- Adaptability in Extreme and Confined Environments
- Designing Assets for Robotic Access
- Simulation, Digital Twins and Multimodality
- Building Trust Between Humans and Machines
- Certification and Who Signs Off on an AI Call
- The Future of Robotics in NDT
- The One Thing Robots Will Never Do
- Key Takeaways
- Frequently Asked Questions
1. Session Overview
Industrial inspection has always relied on something difficult to specify: the accumulated judgement of an experienced technician who can feel when something is wrong. A hand on a transducer that adjusts pressure without being told. An eye that notices the corrosion pattern is subtly different from last year's. This session asked directly whether a drone, a crawler or an automated cell can ever replicate that.
The panel was deliberately assembled from different corners of the automation landscape: confined-space drones, stationary aerospace inspection cells, robotic platforms for hazardous environments, and an NDT division head working primarily with conventional methods in heavy marine fabrication. That spread produced a more honest conversation than a purely vendor-led panel would have, with genuine disagreement about how close the technology actually is.
The session was also notable as the first NDT Talks episode moderated by a woman, with Adriana Nazera of Voliro leading the discussion.
NDT Talks Episode 4 — Robotics, Automation and the Future of Industrial Inspection:
2. Meet the Panel
3. Can a Robot Replicate the Human Touch?
The opening question was put plainly: can an automated system ever develop the sensitivity a human inspector brings to detecting a crack or assessing corrosion? The answers formed a spectrum, and the disagreement between them is the most useful part of this session.
The View From the Field
Mr. Sala Al-Sheikh — NDT Division Head, International Maritime Industries
“For a robot in this field it will not be easy. As a human being we can adjust ourselves according to the standard regulation, and if there is any mistake we can go back and discuss it with expert people. But with a robot, once it gets an error we need to start from the beginning. So it will be a long process. I am always depending on a human being, because the human being is the power of the working. I do not have that much experience with robotics, but I have already worked with automated systems and phased array, and still the human being is the major impact in this field.”
A Different Problem in the Production Cell
Mr. Grafenberger reframed the question from the perspective of stationary automated systems, where the part comes to the machine rather than the machine going to the part. His answer identified a technical problem that is easy to overlook.
Mr. Thomas Grafenberger — Fill GmbH
“All of the parts finally assembled into an aircraft have some deviations from the CAD, from the ideal world, from the nominal world. A human with a transducer in the hand would always move smoothly on a part, no matter whether it deviates from the CAD nominal or not. In the case of an automated system, you need to feed in a program, and that program is mostly based on the nominal. Obviously with these deviations we are creating challenges, and this is exactly where we are investing the majority of our R&D budget: to address this deviation from the digital world, where everything is ideal and you can simulate everything, but in reality it is different.”
The Question Has Changed in Five Years
Mr. Antonio Valerio — Managing Director USA, Flyability
“Our technology by default tends to get deployed in places where humans should not go in the first place, more often than not by legislation. So we have an increased obligation to make sure the technology can substitute or accompany the experience of human experts. If you had asked the same question four or five years ago, the answer would have been a lot more straightforward: no, absolutely, the collection of human experiences an NDT inspector brings to the field is not replaceable. However, with the giant leaps in AI, specifically image defect recognition, it is not out of the realm of possibility that we will be able to condense hundreds of thousands of NDT reports and teach the robots to recognise the difference between different kinds of corrosion and defects.”
Not Yet, But That Is Not the Point
Mr. Kushal Shetty — Business Development Manager, Edify Robotics
“Right now, definitely not. We are not really anywhere close to it. But the reason companies like Flyability and Edify Robotics exist is because there are dangerous environments where people are having to go right now to perform inspections. Even if you cannot get the exact same data set you would get if a human was in that environment, you can still get a data set that leads you to being able to make a decision on these critical assets. What we are seeing more and more is not just being able to grab those data sets, but being able to get repeatable data sets, which means you can make decisions based on the changes. That is something humans somewhat struggle with, because performing the exact same task again and again is quite hard for a human without some variation.”
The reframing that emerged: the useful question is not whether a robot can match a human inspector's tactile judgement. It is whether the data a robot can collect — in places humans cannot safely go, and with a repeatability humans cannot match — supports a sound engineering decision. On that measure, the answer is already yes in many applications.
4. The Sensor Problem: Built for Hands, Not for Robots
Asked how close robotically acquired data is to what a human inspector would obtain with the same method, Mr. Shetty identified what he described as one of his biggest frustrations with the current state of the industry.
Mr. Kushal Shetty — Business Development Manager, Edify Robotics
“I do not think there are that many sensors being developed specifically for robotic systems. You are always trying to incorporate something designed for a human into a robotic platform. That is part of what is missing right now, to an extent where you do not have the ability to get the full range of advanced non-destructive testing capabilities in all robotic platforms. There is a difference between when a human is in a certain place versus a robotic system, how they are adhering to the area, how they are moving in it, and what the form factor of the probe is like. But if it is something compliant with current technology, then you are able to get a very similar data result.”
Mr. Antonio Valerio — Managing Director USA, Flyability
“Right now it is the world of robotics that has had to adapt to the world of NDT, and basically retrofit existing modules onto the robots. That is a given, considering the NDT industry has been around a lot longer than drones. Going forward I think it is going to be interesting to see, not necessarily a reversal, but an adaptation of the NDT industry towards robotics. NDT companies working more closely with robotics providers to make sure the technology fits drones and robots, rather than the other way around.”
This is arguably the most actionable insight in the session for equipment manufacturers. The probes, wedges and couplant arrangements in current use were designed around the ergonomics of a human hand and arm. Form factor, cable management, coupling pressure control and probe mounting geometry all assume a human operator. Purpose-built robotic sensor design remains a largely unaddressed opportunity.
AI-assisted interpretation is already well established in radiography, see our deeper look at how it's being applied to digital X-ray evaluation.
→ Read: AI Digital Radiography Interpretation
5. Using the Process Sensor for Its Own Positioning
Mr. Grafenberger contributed a technical idea that stood out as genuinely novel within the discussion: using the ultrasonic transducer's own returned data to correct its positioning in real time, rather than relying on a separate array of auxiliary sensors.
Mr. Thomas Grafenberger — Fill GmbH
“If you would like to inspect a surface with an automated system, you need feedback to position the sensor accurately on the surface — for UT inspection, normal to the surface at a specific distance. An underestimated potential is utilising the feedback of the process sensor itself for this positioning correction. Typically you see a lot of sensors around the UT sensor on the end-of-arm tool: structured light scanners providing point cloud information, distance sensors, tactile or contactless, even LiDAR, all with the purpose of reconstructing the surface. But the process sensor itself, the UT transducer, is already providing so much data which can be used in real time to adapt the inspection trajectories.”
Why this matters commercially: as Mr. Grafenberger noted, the approach requires no additional hardware. Removing the auxiliary sensor stack from an end-of-arm tool reduces cost, weight, calibration burden and failure modes simultaneously. The constraint is that it suits serial production of similar parts with minor deviations, rather than genuinely unpredictable field geometry.
6. The Data Flood and What to Do With It
A consistent theme across the panel was that robotic inspection has fundamentally changed the volume of data available, and that the industry has not yet caught up with the consequences.
Mr. Antonio Valerio — Managing Director USA, Flyability
“With the use of robots we have basically opened the floodgate for a quantity of data points that is absolutely unprecedented. Compared to what NDT inspectors were previously able to capture in the same time frame, current robotic platforms can increase the number of single data points exponentially. This creates an obvious issue with the review and analysis of these data points, where currently the human eye and human experience is an absolute necessity. Going forward, artificial intelligence can really help to distinguish, in the noise of all the data captured, what is actually relevant, what is valid, what is a false negative, and most importantly how to aggregate this information and distribute it efficiently among the right stakeholders. That is really where a lot of value is currently being lost.”
Mr. Al-Sheikh offered the practical counterweight from a fabrication environment. Automation raises throughput substantially, but when a genuinely ambiguous indication appears, the workflow still routes back to a human.
Mr. Sala Al-Sheikh — NDT Division Head, International Maritime Industries
“For robotics, the good thing is that production will be very high. If they request ten products they can produce it in less time. That is the advantage. But still, once there is some defect and they have doubt in that defect, we go manually and cross-check it to clarify whether it is a real defect or not an indication. With this technology it will be easy for us, it will support us in our job, but still the human being in that part is always the impact.”
Volume as the Trust Argument
Mr. Kushal Shetty — Business Development Manager, Edify Robotics
“There is an undeniable advantage that humans have in terms of dexterity and tactile capabilities. However, there is a significant limit on where we can acquire data and how much data we can acquire. As soon as that larger data set becomes more understood in terms of how much value it can bring, that is where we will see a significant shift. Gecko Robotics is everywhere, everyone knows about them now. They are not necessarily using the most advanced NDT modalities, but they are acquiring so much data and running it through so much contextual analysis using AI that they are able to leverage a lot of benefits out of it. That is going to be the thing that shifts the needle, more than making people believe a robot is actually going to function as well as a human, because frankly it is not, for a very long time.”
Permanently installed sensors are a growing alternative to periodic robotic inspection, see our full guide to IoT-based continuous monitoring.
→ Read: IoT and NDT for Monitoring and Predictive Maintenance
7. Adaptability in Extreme and Confined Environments
Asked how well robotic systems cope with environments that change unpredictably, the panel pointed to a decade of measurable progress.
Mr. Antonio Valerio — Managing Director USA, Flyability
“The level of adaptability is already very significant. Over the last ten years we have been able to adapt to a multitude of incredibly dangerous situations, from nuclear power plant radiation to toxic environments in sewers to very unsafe underground mining environments. It has been fascinating to see how some of the robot dogs have adapted to environments as challenging as offshore platforms. For outdoor drones, in the space of a few years we went from systems that were very difficult to operate in unfavourable weather to systems that can fly in almost all conditions.”
Contextual Awareness as the Enabler
Mr. Kushal Shetty — Business Development Manager, Edify Robotics
“Take Flyability, for example. They have introduced functionality called return to home, which automates the system in a confined space so it can fly back, and if you block off its exit path it relocates and looks for another way out. That is amazing adaptability in an otherwise impossible scenario. Even if a human was piloting it, it would probably be a more difficult task than the robotic system managing to manoeuvre out, just because it has more contextual data. Context is the big thing in all of this. When robots have a great amount of context about the environment they are in, they are able to make better decisions.”
Mr. Shetty was careful to note where the current limit sits. In the nuclear work Edify Robotics performs, a human is still driving the robotic system, aligning the mechanical arm, and verifying that a task such as unbolting is proceeding correctly. Autonomy in navigation is considerably further advanced than autonomy in manipulation.
8. Designing Assets for Robotic Access
One of the most forward-looking points of the session concerned a shift in how new assets are being specified, and it reframes adaptability entirely.
Mr. Kushal Shetty — Business Development Manager, Edify Robotics
“A lot of assets on sites are not necessarily designed for robotic systems to be able to access or inspect easily. But we are being reached out to more and more where people are saying, we have an asset we are going to be delivering in 30 years, for a nuclear plant or for ammonia storage tanks, and we want this to be ready for robotic inspection. We want to design it so that robots can actually access and do the work. That is a fundamental shift, because adaptability is not only about how we make something work where a human can work. It is about adapting the environment as well, to make it work for a robotic system.”
Strategic implication: for asset owners commissioning long-life infrastructure, robotic inspectability is becoming a design requirement rather than a retrofit problem. Access hatches, internal geometry, surface finish, lighting provision and landing or docking points can all be specified at design stage at negligible cost, versus enormous cost once built.
9. Simulation, Digital Twins and Multimodality
Mr. Grafenberger described how heavily the aerospace production side now depends on simulation before any hardware touches a part.
Mr. Thomas Grafenberger — Fill GmbH
“In the meanwhile it is standard that you simulate all these inspections prior to occupying the asset on the shop floor with a real part. You can parallelise the entire commissioning of such a system, and also the introduction of new parts, on a digital twin. You can predict cycle times. You can even simulate how the sound beams of the UT propagate into the specimen with specific tools. If I compare the share of digital simulation in such a system now versus ten years ago, the share of just digital input on a project, simulation and programming tools to assist the operators, has significantly grown. A system like the one behind me could not even be commissioned on the shop floor without sufficient simulation in the digital world beforehand.”
Changing the Method, Not Just the Manipulator
Mr. Grafenberger also made a point that reframes automation strategy: sometimes the right answer is not to automate the existing method, but to replace it with one better suited to automation.
Mr. Thomas Grafenberger — Fill GmbH
“For some specific applications it is possible to think about just replacing the legacy inspection technology, which in this case is water-coupled UT, with a new alternative method that has different requirements. We can inspect this kind of part with laser-excited UT, so we can get rid of the coupling water, which helps in some environments to reduce risks and ensure accessibility. Sometimes parts in aerospace should not get wet for inspection, so nowadays those parts are sealed, which is a very challenging and time-consuming manual task. We call it multimodality. Our inspection systems are not necessarily focused on ultrasonics, rather than having the entire palette of inspection technologies which we can exchange fully automatically — UT, laser-excited UT, thermography, X-ray — and all these modalities combined in one automated system.”
The certification pathway for methods like laser-excited UT follows a similar Level I/II/III framework, see our full breakdown of ultrasonic testing standards and codes.
→ Read: Ultrasonic Testing Standards and Codes
10. Building Trust Between Humans and Machines
The moderator raised a question not in the prepared set but which the discussion had made unavoidable: how does the industry build trust in systems that will inevitably make mistakes?
Mr. Antonio Valerio — Managing Director USA, Flyability
“It boils down to three things: acknowledge the mistakes, listen, and improve. There is a quote from Beckett that comes to mind: try again, fail again, fail better. The burden of proof is always on the innovators. We deal with people naturally embedded in a risk-averse environment — that is their bread and butter, that is their job. For our technology to be fully accepted it has to show an error level at least as low as current methods. Otherwise it is very difficult for operations managers and risk managers to accept the change. The important thing is to realise that every single product is a prototype for the next generation.”
The Honest View From a Prospective Adopter
Mr. Al-Sheikh's position is worth quoting because it represents a very large segment of the industry that panels like this rarely hear from directly: interested, not yet convinced, and constrained by cost and organisational inertia rather than by technical scepticism.
Mr. Sala Al-Sheikh — NDT Division Head, International Maritime Industries
“It will be a very huge impact, very huge things for the future, because it will save many things for our product. But still, because it is in the first stage, it needs our trust. So let us see, let us try. I would like to try this thing. In our company it is still under discussion, because the cost, the timing, everything, for them this is the main thing.”
He also described a concrete automation attempt that had not gone smoothly: using bots to enter inspection data into a reporting system. The obstacle was not the inspection at all, but the inventory data — tool and equipment serial numbers, naming conventions, equipment types — which proved difficult to hand to an automated process reliably. It is a useful reminder that the friction in NDT automation is frequently in the surrounding data infrastructure rather than in the measurement itself.
11. Certification and Who Signs Off on an AI Call
An audience question from Roberto Passerini asked about the regulatory pathway for certifying a transition from manual NDT methods to robotic and automated systems for aerospace materials. Mr. Grafenberger addressed it directly.
Mr. Thomas Grafenberger — Fill GmbH
“At the moment there is the well-known concept of Level I, Level II and Level III. The Level III is ultimately allowed to sign off a procedure. A Level II is allowed to sign off that an inspection was done properly and evaluated properly. This topic is more and more becoming a topic in the NDT aerospace community, because the big question is: if you are utilising AI to evaluate parts, who at the end is responsible for a false negative or false positive? Nowadays, even if you use the best algorithms to bring some decisions, at the end a human must sign off this result. This is the status quo. But there are many discussions ongoing about how this might change in the future.”
Why UT Is Harder Than X-ray for AI
Mr. Thomas Grafenberger — Fill GmbH
“AI-assisted evaluation in X-ray has been state of the art for a long time, because it is just vision processing. But in UT it is a bit different, especially when talking about composite materials, which are very anisotropic. It is even for an AI quite challenging to find some indications in an automated way. The research is at a very early stage on this topic.”
The moderator noted that aerospace is conservative by design and progresses slowly, and Mr. Grafenberger framed his own expectations in decades rather than years. For anyone building a business case around automated NDT in regulated aerospace applications, that timeline is the realistic planning assumption.
12. The Future of Robotics in NDT
Mr. Thomas Grafenberger — Fill GmbH
“Definitely the future is with robots, and robots are becoming more and more intelligent. But I am of the opinion that we should not try to replicate the human. Rather, we should identify the pain points in the inspection and find the best approach to solve those particular challenges. I have seen several approaches trying to use humanoid robots. I prefer the approach of analysing the pain point and finding the best and appropriate technical solution, rather than trying to replicate it in the way a human would do it.”
Mr. Kushal Shetty — Business Development Manager, Edify Robotics
“There is potential for a significant shift away from the current methodology. Look at installed sensors and how widely they are being adopted now — instruments giving continuous or timed feedback, which gives another set of insight into active assets. We have seen it in pipelines where they have fibre optics or other sensors running along the pipeline so they can tell if there is damage or rupture or bends. Yes, there is definitely a place for NDT deployed by robotics and it is going to grow. But I also think there is going to be a huge adoption of other types of sensing, where you do not necessarily need to be interfering in the environment to get your data back.”
A note on that second point: permanently installed monitoring sensors represent a genuine strategic alternative to robotic deployment, not merely a complement to it. If a sensor is already on the asset reporting continuously, neither a human nor a robot needs to travel to it. This is a different competitive axis from the human-versus-robot framing that opened the session.
13. The One Thing Robots Will Never Do
The closing question asked what task robots will never take over from humans in industrial inspection. The answer came back inverted, and it landed as the strongest note of the session.
Mr. Antonio Valerio — Managing Director USA, Flyability
“Getting injured. Which is why we should really aim at using robots as much as possible. A robot can get damaged, but it cannot get injured. Somebody was mentioning in the chat whether robot dogs come with floating devices on offshore platforms. They do not need to. At the end of the day, if they sink, they sink, and insurance is going to take care of that. We often say in this industry that you cannot put a price on human life. Turns out you can — insurance companies certainly do it, shareholders certainly do it. So the thing in the inspection world that robots can never do is get injured.”
The moderator added a line she had heard the previous week at a commercial UAV trade show, which the panel let stand as the closing thought: an inspector had observed that he has attended the funerals of many colleagues in this industry, but he has never attended a drone funeral.
14. Key Takeaways
- No panelist claimed robotic systems currently match human tactile judgement. The consensus is that they do not, and will not for some time.
- The more useful question is not equivalence but sufficiency: can robotically acquired data support a sound engineering decision on a critical asset? In many applications it already can.
- Robots decisively outperform humans on repeatability, which enables trend-based decision-making across inspection cycles in a way manual inspection struggles to support.
- Almost no NDT sensors are designed specifically for robotic deployment. Current practice retrofits human-ergonomic probes onto robotic platforms, and this is a significant unaddressed constraint.
- Using the UT transducer's own returned signal for real-time trajectory correction can eliminate the auxiliary sensor stack entirely, reducing cost and complexity.
- Robotic deployment has increased available data volume by orders of magnitude, but analysis, aggregation and distribution of that data lag badly behind acquisition.
- Automated navigation in confined and hazardous spaces is now genuinely capable. Automated manipulation still requires a human in the loop.
- Asset owners are beginning to specify robotic inspectability at design stage for long-life infrastructure, which is a more effective route than retrofitting access.
- Simulation and digital twins are now standard practice in automated aerospace inspection, with commissioning effectively impossible without them.
- Sometimes the right automation strategy is to replace the method rather than automate it, for example laser-excited UT eliminating water coupling.
- A human must currently sign off on any AI-assisted evaluation in regulated aerospace work. AI-assisted X-ray evaluation is mature; AI for UT on anisotropic composites is at an early research stage.
- Trust is built by acknowledging error, incorporating feedback, and demonstrating an error rate at least as low as existing methods.
- The strongest argument for robotic inspection is not data quality. It is that robots cannot be injured.
15. Frequently Asked Questions
What is robotic NDT inspection?
Robotic NDT inspection is the deployment of non-destructive testing sensors on an automated platform rather than by a human operator. Platforms include confined-space drones, crawlers, robot dogs, magnetic wall-climbing units, and stationary robotic cells in production environments. The sensors carried are largely the same as those used manually — ultrasonic transducers, eddy current probes, visual and thermal cameras — though probe form factor and mounting are adapted for robotic carriage.
Can robots replace human NDT inspectors?
Not currently, and the panel was consistent on this. Robots cannot yet replicate the adaptive tactile judgement of an experienced inspector, and ambiguous indications still route back to a human for verification. Where robots do replace humans is in environments too dangerous or legally restricted for human entry, and in tasks requiring high repeatability across inspection cycles. The realistic near-term model is collaboration rather than replacement.
What are the main limitations of automated NDT?
The panel identified several: sensors designed for human ergonomics rather than robotic deployment; difficulty accommodating deviation between the nominal CAD geometry and the as-built part; limited autonomy in manipulation as opposed to navigation; assets not designed for robotic access; the volume of data outpacing the industry's ability to analyse it; and regulatory frameworks that still require a human signature on every evaluation.
How are drones used in confined space inspection?
Collision-tolerant drones are flown into vessels, tanks, boilers, sewers, mine workings and other confined spaces to capture visual and, increasingly, ultrasonic thickness data without requiring human entry. This removes the need for scaffolding, confined-space entry permits and rescue standby, and eliminates exposure to atmospheres that are toxic, oxygen-deficient or radiologically hazardous. Modern systems include autonomous return-to-home behaviour that will find an alternative exit route if the original path becomes blocked.
Is AI reliable enough for NDT defect recognition?
It depends heavily on the modality. AI-assisted evaluation of X-ray images is mature and long established, because it is fundamentally a vision-processing task. Ultrasonic evaluation is substantially harder, particularly on anisotropic materials such as composites, and research there remains early. Across all modalities in regulated industries, a qualified human must currently sign off the final result regardless of what the algorithm reports.
Should new assets be designed for robotic inspection?
Increasingly, yes. Panelists reported growing demand from owners of long-life assets — nuclear facilities, ammonia storage, major infrastructure — specifying robotic inspectability at design stage. Access geometry, internal clearances, surface condition and docking provision can be built in at negligible cost during design, and at very high cost or not at all afterwards.
Where can I watch this NDT Talks session?
The full session is available on the OneStop NDT YouTube channel:
Ads
