Data Center Robotics: Who Checks the Robot's Work?

• Robots are starting to do physical repair work in data centers, mostly as supervised pilots at the partial and high levels of automation.
• The party that does the work should not certify it, the rule behind independent commissioning agents.
• Moving cables during a repair can cause transient packet loss in the cables touched, so a check must cover the area around the job.
• On a data center cabling test board, the best learned robot policy in a 2026 benchmark succeeded 78% of the time; real racks add variety.
• As automation rises from Level 0 to Level 4, fewer people watch each job and the independent record has to carry more.
• New crews are usually checked by a senior technician walking the row, the scarcest person on the build.
• One standard, applied the same way to every site, crew, vendor and robot, turns drift into data.

Short answer: Yes, in early form. Microsoft Research has built prototype robots that reseat and clean transceivers, Google has trialed robots for handling hard drives and equipment, and most of this work still runs as supervised pilots.

The work robots are reaching for is the repetitive physical work that keeps a hall running. Microsoft Research's HotNets 2024 paper describes prototypes that reseat a transceiver by pulling it, waiting a few seconds and pushing it back, and that detach a fiber, inspect and clean its end face and reconnect it.

The motivation is speed. The paper notes that a physical repair today runs on a timescale of days, with some high priority repairs done in hours, and that gray or transient failures are common, so faster physical work means less time with a link out of service.

Other operators are moving too. DCD reported in 2024 that Google has trialed robots for managing hard drives and other equipment, and DCD's coverage of the Microsoft work describes an early stage, modular approach.

Most of this is still supervised, and the Microsoft team places its own work between the partial and high levels of automation. For what robots can realistically take on today, see our guide to whether robots will do routine data center maintenance.

Robots are starting to do the work. The question is who checks it.

Short answer: Because the party that does the work cannot objectively certify it. Construction learned this long ago, which is why owners hire an independent commissioning agent, and the same logic applies to a robot, its vendor and a crew checking itself.

Commissioning guidance makes independence the point. The ACG commissioning guideline describes the commissioning authority as independent of the designers, contractors, vendors and suppliers, because contractors find it hard to test and assess their own work objectively.

The technicians, contractors and commissioning engineers we interviewed said the same thing in plainer words. Owners hire an independent agent because the contractor cannot verify its own work, and positive confirmation, where each item is recorded as checked and passed, beats a report where silence counts as a pass.

A robot is a contractor with a very consistent hand. It may do the same task the same way a thousand times, but if its own sensors decide whether the task succeeded, any blind spot in those sensors becomes a blind spot in the result.

The same holds for the company that sells the robot. A vendor grading its own robot's work has the same conflict as a builder certifying its own building, however good the robot is.

The party that does the work should never be the one that signs it off.

Short answer: Whatever it was not looking at. A robot checking its own task looks at the task, not the rack around it, and moving cables during a repair can disturb the ones next to the job.

Physical work has side effects. Microsoft's paper notes that technicians moving fiber cables to reach a component can cause transient packet loss in the cables they touch, and that motion near hardware creates vibrations that can cause further failures.

The paper's own transceiver robot is designed to minimize accidental contact with nearby cables, which tells you the risk is real. A robot that verifies its own job confirms the transceiver it reseated, not the cord it brushed on the way in.

What a doer cannot see about its own work

• Its own mistakes: any blind spot in the sensors that judged the task.
• Collateral disturbance: the cords, caps and connectors next to the job.
• Other jobs: work done by other teams, other vendors and other robots in the same rack.
• Drift: how the same job is done differently between sites, shifts and crews.

The fix is a check of the whole area after the work, not only the job. Our guide to intermittent fiber link failures shows how a link disturbed by nearby work can look healthy for days before it starts to flap.

Checking the job is not the same as checking what the job touched.

Short answer: Improving fast, and best at specific, well defined tasks. On a data center cabling test board, the best learned policy in a 2026 benchmark grasped, cleaned and inserted a single cable 78% of the time across 48 trials.

The Industrial Dexterity Benchmark, published in July 2026, tested robot policies on a data center cable manipulation board. The best setup, using several camera views, reached a 78% combined grasp and insert success rate across 48 trials, against 36% for a single camera baseline.

That is real progress on a bench built for the task. Real racks add variety: different cable types, tighter spaces, other cables in the way and lighting that changes from aisle to aisle.

Variety is what makes manipulation hard. In THE COLOSSEUM benchmark (RSS 2024), robot success fell by 30% to 50% when a single factor in the scene changed, and by 75% or more when several changed at once.

Measuring this well is still an open problem. NIST's robotics program points to a lack of metrics, benchmarks and test methods as a barrier to verifying that robotic systems meet their requirements.

Our view follows from the numbers. Specific actions are far more reliable than general ones, so the harder and more varied the work, the more it needs an independent check.

None of this is a reason to wait. Supervised robots already save technicians walks and lifts, and every gain in what robots can do raises the stakes of checking what they did.

Robots will keep getting better. The check matters more as they do more.

Short answer: Microsoft Research describes five levels, from no automation to full automation. The higher the level, the less a person watches each job, so the more an independent record has to carry.

The levels come from the HotNets 2024 paper. The checking line under each one is ours, and it shows how verification has to change as people step back.

Five levels of repair automation, and who checks each

1. 1. Level 0, no automation: a technician does the repair, and a senior technician checks a sample.
2. 2. Level 1, operator assistance: tools guide the technician, and the check is still a person's.
3. 3. Level 2, partial automation: a robot does specific tasks under supervision or remote control, and the supervisor watches the task, not the rack around it.
4. 4. Level 3, high automation: a robot does whole tasks with limited supervision, and an independent check of the job and its neighbors becomes the main safeguard.
5. 5. Level 4, full automation: nobody watches the job, and an independent record is the only proof the work was right.

Most operators will run several levels at once for years. A hall may have technicians, supervised robots and a few fully automated tasks side by side, which is exactly why the check needs to be the same for all of them.

The distinction between doing and checking matters here. Our guide to monitoring, inspection and robotic action explains why the three are different jobs, and why the one that acts should not be the only one that looks.

The less anyone watches the work, the more the record has to carry.

Short answer: Today, mostly a senior technician walking the row, the scarcest person on the build. As thousands of new technicians join the industry, an independent check takes that walk off the senior people so they can go back to the work only they can do.

The workforce is growing fast, and it has to. CSIS estimates that AI data centers could need roughly 140,000 extra skilled trades workers by 2030 in its high case.

Training programs are scaling to meet it. Large operators and their partners are training thousands of new fiber technicians, and Microsoft's Datacenter Academy reports more than 12,522 students enrolled.

New crews make the mistakes every new crew makes. The technicians, contractors and commissioning engineers we interviewed told us that what usually catches them is a senior technician walking the row, and that senior technician is the person every project is short of.

An independent check changes who spends time where. The new crew gets a punch list with a photo for every item while still on site, and the senior technician reviews the exceptions instead of walking every row, as our guide to reducing fiber commissioning rework describes.

Crews also change from site to site. A contractor that does excellent work in one hall may send a different crew to the next, so the check has to follow the work, not the reputation.

New crews do not need someone standing over them. They need a record of what they got right and what to fix.

Short answer: Because standards drift between sites, contractors, shifts and machines. One check applied the same way everywhere shows which crews, vendors, cable batches and robots produce defects, and replaces hours of video nobody has time to watch.

Every site writes down its standards for cabling, labeling, dressing and safety, and every site applies them a little differently. Inspectors drift too: two people walking the same row on different days will not record the same things.

One check, applied the same way everywhere, turns that drift into data. It produces deviation rates per crew and per vendor, and a scorecard of which installers, crews, cable batches and parts produce defects, compared site to site.

What one standard gives you

• No inspector drift: the same checks, the same way, on every site and shift.
• Deviation rates: per crew and per vendor, so help goes where it is needed.
• A vendor scorecard: installers, crews, cable batches and parts compared site to site.
• One record across a mixed fleet: the work of every robot and every crew in one format.
• No video to watch: a verified record of exactly what changed, instead of hours of footage.

The same applies beyond fiber. Power paths and cooling connections need the same independent record, as our guides to A and B power feeds and liquid cooling leak detection explain.

One standard is only a standard if the same check enforces it everywhere.

Short answer: Each item checked, with a photo, a location and a time, before and after images of what changed, and clear rules: unreadable items go to a person, closed doors are recorded as not inspected, and nothing is reported clear unless every point was seen.

The independent verification test

1. 1. Did someone other than the doer confirm the work?
2. 2. Was the area around the job checked, not only the job?
3. 3. Is there a record for each item, with a photo, a location and a time?
4. 4. Can unreadable or not seen ever count as a pass?
5. 5. Is the same standard applied on every site, crew, vendor and robot?

If the answer to the fourth question is yes, the record cannot be trusted. Everything else follows from evidence rules that are written down and applied the same way every time, the principle behind a verified inspection record.

A good record also shows position, not just a photo. Images taken from the same position and angle every time can be compared round to round, so a change shows up as a change.

Keep the record where the owner can use it. A finding that lives in a vendor's portal or on a crew's phone is hard to compare across sites, while a record the owner holds, in one format, is what makes the scorecard possible.

A record you can trust says what was checked, what was seen and what was not.

Short answer: A FacilityOps robot walks the row after any robot or crew has worked in it and records what changed, with before and after images, a photo for every item and the same evidence rules every time. We check the work. We never do the work.

FacilityOps AI (RoboSkill Inc.) builds an autonomous inspection platform for data centers. Somebody has to check the robots, and the crews, and that is the job we built it for.

Our inspection robots run fully autonomously, and a technician can take manual control at any time. The platform is hardware agnostic: inspection robots and sensor hubs are the most common way we capture data, but it works with any capture source.

What the robot checks

• What changed: a FacilityOps robot walks the row after any robot or crew has worked in it and records what moved, appeared or disappeared, with before and after images in our 3D digital twin.
• Disturbance and attribution: after anyone works in a rack, the next round shows what was disturbed nearby and when.
• Positive confirmation: every port and item in scope recorded as checked, with a photo and its location, as in our guide to automated fiber optic inspection.
• One standard: the same checks on every site, crew and vendor, with deviation rates per crew and vendor and a vendor scorecard.
• One record: the work of a mixed fleet and every crew in one format, instead of hours of video.

Built to be trusted

• We check the work. We never do the work. No plugging, unplugging, reseating, cleaning, part swaps or repair, in any phase, so we never grade our own work.
• People stay in the loop: anything the robot cannot read clearly goes to a person, a closed door is recorded as not inspected, nothing is reported clear unless every point was seen, and a person reviews every finding before it reaches you.
• Everything runs on site: detection and the 3D model run on the robot's onboard computer, and you own every image and record. Patent pending.
• On our roadmap: issuing an independent completion record for work done by other robots, as it happens.

Where it stands today

Whole facility rounds, the 3D digital twin and change detection run in a production data center today. Since July 28, 2026, a FacilityOps robot has run inspection rounds at the CRSC Data Center, powered by Solutions IS, in Houston, completing 19 autonomous inspection runs and more than 1,600 observations in its first three weeks.

A one row trial shows you the record after your own crews or robots have worked, judged on your own data. It takes the checking walk off your senior technicians, so they spend their time on the work only they can do.

• Repair automation level: how much of a repair a machine does without a person, from Level 0 to Level 4.
• Teleoperation: a person controlling a robot remotely.
• Reseating: pulling a part such as a transceiver and pushing it back in.
• Collateral disturbance: damage or movement caused next to a job, not by the job itself.
• Gray failure: a partial fault that health checks miss while workloads suffer.
• Commissioning agent: the independent party who verifies and signs off a build for the owner.
• Independent verification: a check of the work by a party that did not do it.
• Positive confirmation: a record that each item was checked and passed, rather than silence counting as a pass.
• Change detection: comparing one round with the last to see what moved, appeared or disappeared.
• Inspector drift: the way different people, shifts and sites record the same check differently.
• Vendor scorecard: defect rates compared across installers, crews, cable batches and parts.
• Manipulation: a robot grasping, moving or inserting objects with an arm and gripper.

What tasks can robots do in a data center today?

Mostly specific, repetitive physical tasks in supervised pilots: reseating and cleaning transceivers, handling hard drives, moving equipment and running inspection rounds. Inspection robots are the most mature, because they look rather than touch. Robots that do repairs are improving fast but still work mostly at the partial or high automation levels, with people supervising.

Will robots replace data center technicians?

No. The industry needs more skilled people, not fewer, and CSIS estimates AI data centers could need roughly 140,000 extra skilled trades workers by 2030 in its high case. Robots take on repetitive work and checking walks, which frees technicians for work that needs judgment, and someone still has to verify what the robots do.

Can a robot reseat or clean a fiber transceiver?

Prototypes can. Microsoft Research has described robots that reseat a transceiver by pulling it, waiting and pushing it back, and that detach a fiber, inspect and clean its end face and reconnect it. These are research systems, and the same paper notes that moving cables can cause transient packet loss in the cables touched.

How reliable are data center repair robots?

Improving, and best at specific tasks. In a 2026 benchmark on a data center cabling test board, the best learned policy grasped, cleaned and inserted a single cable 78% of the time across 48 trials. Real racks add variety, and other research shows success falls sharply when the scene changes, so independent checking matters.

What are the levels of data center repair automation?

Microsoft Research describes five: Level 0, no automation; Level 1, operator assistance; Level 2, partial automation, where robots do specific tasks under supervision or remote control; Level 3, high automation, with limited supervision; and Level 4, full automation. The higher the level, the more an independent record has to carry.

Who is responsible if a robot causes an outage?

That depends on the contracts between the operator, the robot vendor and any service provider, so settle it before robots go live. Whoever is responsible, the fastest way to establish what happened is an independent record of the rack before and after the work, showing what changed, when and what was disturbed nearby.

How do you verify a robot completed a repair correctly?

Check the work with something other than the robot that did it. Confirm the repaired item, check the area around it for disturbance, record each item with a photo, a location and a time, compare before and after images, and confirm the result against link or system telemetry once it is back in service.

Can moving cables during a repair cause network problems?

Yes. Microsoft Research notes that moving fiber cables to reach a component can cause transient packet loss in the cables touched, and that motion near hardware can create vibrations that cause further failures. That is why a check after any repair should cover the cables and connectors around the job, not only the job.

Why must a commissioning agent be independent?

Because it is hard for anyone to test their own work objectively. Commissioning guidance describes the agent as independent of designers, contractors, vendors and suppliers, so the party that installed the equipment is not the one certifying it. The same principle applies when a robot or its vendor does the work.

Is there a shortage of data center technicians?

Yes. CSIS estimates that AI data centers could need roughly 140,000 extra skilled trades workers by 2030 in its high case, and large operators and their partners are training thousands of new fiber technicians. Senior technicians, who usually catch new crews' mistakes, are the scarcest people on most builds.

How do you check the work of newly trained technicians?

Check every item after the crew finishes, while they are still on site, and give them a punch list with a photo and a location for each defect. Track defect rates per crew so help goes where it is needed. That way the senior technician reviews the exceptions instead of walking every row.

Can one inspection platform check work done by different robots?

Yes. A FacilityOps robot records what changed after any robot or crew works in a rack, with before and after images and the same evidence rules every time, so the work of a mixed fleet ends up in one record. It never does the work itself, so it never grades its own repairs.

↑ Back to contents

Robots and new crews will do more and more of the physical work in data centers, and that is good news for an industry short of hands. But the party that does the work should not be the one that signs it off, whether it is a contractor, a robot or a vendor.

An independent record, with one standard, positive confirmation and a check of everything the work touched, is what lets operators trust the work at scale. It is also what lets them hand off the checking walk.

The more the work is automated, the more the record has to carry.

• Hong, Rowstron and colleagues, Microsoft Research: Self maintaining networked systems: the rise of datacenter robotics, ACM HotNets 2024
• DCD: Microsoft Research details early stage modular robotics for data centers, November 5, 2024
• DCD: Google tests robots for managing data center hard drives and equipment, October 19, 2024
• He and colleagues: Industrial Dexterity Benchmark, July 2026
• Pumacay, Fox and colleagues: THE COLOSSEUM: a benchmark for evaluating generalization for robotic manipulation, RSS 2024
• NIST: Robotic Systems for Smart Manufacturing Program, updated March 26, 2025
• Girishankar and Smith, CSIS: GenAI's human infrastructure challenge, September 16, 2025
• Microsoft: Microsoft Datacenter Academy helps launch tech careers, September 21, 2026
• AABC Commissioning Group: ACG Commissioning Guideline, 2005
• Field figures: FacilityOps interviews with data center technicians, contractors and commissioning engineers, October 2026

• Will Robots Do Routine Data Center Maintenance?
• Automated Fiber Optic Inspection: How Robots Check Every Port, Connector and Label
• Data Center Commissioning Checklist: What to Inspect Before Handover, Trade by Trade