The Numbers Say Robots Are Safe. The Numbers Are Incomplete.

There are more robots working alongside people than at any point in human history, and we know remarkably little about what happens when they get it wrong.

Both halves of that sentence are true at once, and the gap between them is one of the most significant and least discussed problems in industrial safety today.

Start with the scale

The number of industrial robots in operational use worldwide reached 4,664,000 units in 2024, a 9 percent increase over the previous year. In that single year, 542,000 new robots were installed, more than double the annual number a decade earlier.¹ This is not a gradual trend. It is a rapid acceleration.

The warehouse is where this shows up most vividly. Amazon now operates more than one million robots across its fulfillment network, a fleet that has grown from roughly 750,000 in 2023 and is closing in on the size of its human operations workforce of about 1.2 million.² The company reports that around 75 percent of its global deliveries now get some assistance from robots.³ The practical result is a working floor where robots and people are approaching a one-to-one ratio, sharing the same space, at the same time, all day.

Robot density tells the same story from another angle. South Korea now records the world's highest density at 1,220 robots for every 10,000 manufacturing employees, with Singapore and Germany not far behind.⁴ Whatever else is true, robots and humans now occupy the same physical environment at a scale we have never attempted before, and one that is increasing rapidly.

Now look at what we actually measure

Given that scale, you would expect a correspondingly large and detailed body of safety data. There isn't one.

A peer-reviewed analysis of OSHA's Severe Injury Reports identified just 77 automation-related accidents in the United States across the eight years from 2015 to 2022. Of those, 54 involved stationary robots and 23 involved mobile robots, together producing injuries, concentrated in amputations, fractures, and similar serious harm.⁵ On the most severe end, a NIOSH analysis identified 41 robot-related fatalities in the United States across the twenty-five years from 1992 to 2017. Within that set, 78 percent occurred in manufacturing.⁶

It is worth noting what sits behind those counts. OSHA's own 2024 injury summary acknowledges there are currently no specific OSHA standards for the use of robotics in workplaces, only general provisions such as lockout and machine guarding applied by analogy.⁷ The measurement is thin because the framework underneath it was never built for robots in the first place.

Set those numbers next to 4.66 million operational robots and the mismatch is impossible to ignore. Seventy-seven recorded accidents across eight years, against millions of machines in daily contact with people, does not read like a complete safety picture. Quite the opposite, in fact.

This is the heart of the problem. Low recorded incident counts are easy to mistake for strong safety performance. But an absence of data is not the same as an absence incidents and subsequent harm. When the surveillance is thin, inconsistent, and voluntary, the low numbers may be telling you more about the reporting system than about the robots working alongside humans in a collaborative environment.

What the counts miss

The individual cases that do surface are a useful reminder of what the aggregate numbers leave out, and where the harm concentrates. In November 2023, an industrial robot at a vegetable packaging plant in South Korea grabbed a worker and pressed him against a conveyor belt, causing fatal head and chest injuries. He was not a line operator. He was there to inspect whether the machine was working correctly.⁸ The same reporting noted it was not isolated: a separate robot seriously injured a worker inspecting an auto parts machine earlier that year, and another had fatally crushed a worker at a dairy plant the year before.

The pattern repeats across geographies. In September 2024, a worker was crushed to death by a robotic machine at a frozen food plant in Wisconsin.⁹ And in a case now working through the California courts, a robotics technician alleges that during a maintenance task at an automotive plant a robot arm activated without warning and struck him with the force of an 8,000-pound counterbalance, pinning him against a conveyor.¹⁰ The common thread running through these is not random malfunction on a busy line. It is a person deliberately inside the machine's working space, doing exactly the maintenance and inspection work the aggregate counts are least equipped to capture.

What happens when reporting is actually required

There is one corner of the autonomous systems world where incident reporting is mandatory rather than incidental: automated vehicles. Since 2021, the National Highway Traffic Safety Administration has required manufacturers and operators to report crashes involving automated driving systems under its Standing General Order.¹¹

The instructive part is what happened once that requirement existed. In the reporting system's first months, NHTSA had already logged 130 automated-driving-system crashes, a volume its previous patchwork of media reports and voluntary disclosures had never surfaced.¹² And the count did not plateau. By 2024, NHTSA was recording 544 crashes involving driverless systems in a single year, an average of roughly 1.5 every day, and the cumulative dataset of automated and driver-assistance crashes had passed 5,000 reports.¹³

The lesson is not that automated vehicles are uniquely dangerous. In fact, because minor incidents are captured more completely under a mandate, straightforward crash-per-mile comparisons tend to overstate automated-vehicle risk rather than understate it.¹⁴ That is the whole point. When you require the data, the incidents appear. Mandated reporting does not create risk; it reveals risk that was always there but previously went uncounted. That single insight reframes every low incident number in the rest of the industry.

Even recorded numbers can mislead

If thin reporting is one failure mode, a second is more uncomfortable: numbers that exist but do not disclose the whole truth of the matter. A 2024 congressional investigation into warehouse safety at the largest operator of workplace robots concluded that the company had understated the danger of its facilities, finding its warehouses recorded roughly 30 percent more injuries in 2023 than the warehousing industry average.¹⁵ The investigation traced part of the risk directly to automation: an internal study of workers picking from robotic shelf units found that injury likelihood rose with the pace of picking, estimating an upper bound of 1,940 repetitive movements in a ten-hour shift.¹⁶ A corporate-wide settlement with the Department of Labor followed.¹⁷

Whatever one makes of a single company's practices, the structural lesson is hard to avoid. A recorded injury rate is only as honest as the system that produces it. "The numbers look good" and "the workplace is safe" are not the same statement, and the distance between them is exactly where after-the-fact measurement fails.

The nuance that keeps this honest

None of this means robots make workplaces more dangerous. The evidence points the other way for routine harm. Research examining greater robot exposure in the United States and Germany found it associated with an overall reduction in the rate of work-related injuries.¹⁸ Automation genuinely removes people from many repetitive and hazardous tasks, and the routine-injury numbers reflect that.

But the same body of work found that the reduction did not extend to the most severe injuries. The routine-harm rate falls while the severe-injury tail persists, and much of the remaining risk concentrates in a specific place: human-robot interaction inside the working envelope, frequently during maintenance, when a person deliberately enters the space a robot occupies.¹⁹ That is precisely the interaction that is hardest to see in sparse, backward-looking accident data, and precisely where the named cases above keep landing.

Why this is a system problem, not a data problem

It would be easy to conclude that the answer is simply more reporting. Better databases would help. But the deeper issue is that safety outcomes measured after the fact, through whatever incidents happen to surface, are the weakest possible form of safety assurance. By the time an incident is counted, the harm has already occurred. That said, this does add credence to the old saying, “what gets measured, gets managed.”

The stronger approach is to build the evidence before deployment, not to wait for it to accumulate afterward. A rigorous functional safety process generates structured, traceable evidence at every stage: hazards identified systematically rather than discovered through injury, safety functions defined and verified against recognized standards, and validation documented in a form an assessor can examine. That evidence exists whether or not an incident is ever recorded, because its purpose is to prevent the incident in the first place.

The deployment numbers will keep climbing. Four and a half million robots will become five, and then more, and the share of them working in close proximity to people will grow with every quarter. The question is not whether we can count the harm accurately after it happens. The question is whether the safety evidence is built in from the start and designed into the system overall, so that the count stays low for the right reason.

That is the difference between a robot that appears safe because nothing has been recorded, and a robot that is safe because its safety was engineered, documented, and proven before it ever shared a floor with a human being. At Fennec, we build tools for the second camp, and we’re ready to help more companies stop measuring safety by what hasn’t gone wrong, and start proving why it won’t.

_Notes

_{1. International Federation of Robotics, World Robotics 2025: Global Robot Demand in Factories Doubles Over 10 Years (ifr.org).}

_{2. Silicon Canals, Amazon now runs more than a million warehouse robots.}

_{3. Ibid.}

_{4. International Federation of Robotics, Robot Density Surges in Europe, Asia, and the Americas (World Robotics 2025).}

_{5. Sanders, Sener and Chen, Robot-related injuries in the workplace: An analysis of OSHA Severe Injury Reports, Applied Ergonomics (2024). PubMed.}

_{6. NIOSH, CDC Center for Occupational Robotics Research; Layne (2023). cdc.gov/niosh.}

_{7. OSHA, 2024 Work-Related Injury and Illness Summary (osha.gov).}

_{8. Associated Press, via CBS News, Industrial robot crushes worker to death as he checks whether it was working properly (November 2023).}

_{9. Reporting on a September 2024 fatality at a frozen food plant in Wisconsin, via Yahoo News / Associated Press.}

_{10. Hinterdobler v. Tesla, filed 2025, Alameda County Superior Court, as reported by CBS 8. Allegations in a civil complaint are unproven.}

_{11. NHTSA, Standing General Order on Crash Reporting (nhtsa.gov).}

_{12. NHTSA, Summary Report: Standing General Order on Crash Reporting for Automated Driving Systems (DOT HS 813 324, June 2022).}

_{13. NHTSA Standing General Order data as reported by Arash Law (544 crashes in 2024) and Craft Law Firm (cumulative ADS and ADAS reports through 2025).}

_{14. On reporting completeness and crash-rate comparisons, see analysis via Kisling, Nestico and Redick.}

_{15. U.S. Senate Committee on Health, Education, Labor and Pensions, The Injury-Productivity Trade-off (Majority Staff Report, December 2024); coverage via NPR.}

_{16. Ibid. (internal study referenced in the Senate HELP Committee report), coverage via KING 5 / Associated Press.}

_{17. U.S. Department of Labor corporate-wide settlement, December 2024, as summarized in reporting on warehouse robotics safety.}

_{18. Robot exposure and workplace injury analysis, United States and Germany. NBER.}

_{19. Winfield et al., Robot Accident Investigation. arxiv.org.}