Do warehouse robots make workers safer? What the data says

Warehouse robots are often sold as a safety improvement: machines do the walking, lifting and pushing, so people get hurt less. The evidence supports part of that claim. The most rigorous study so far found that robotics cut severe injuries at Amazon warehouses but increased less severe ones, and the authors link the increase partly to a faster pace of work. This article sets out what the main studies and reports found, why they differ, and how to measure the effect in your own pilot.

What Amazon says

Amazon says it passed one million robots in its operations in 2025 and describes its fleet as designed to make employees' jobs easier and safer ^[1]. It says its Proteus robot works alongside employees in open, unrestricted areas of its sites ^[1]. On safety outcomes, Amazon says company data shows that in 2022 recordable incident rates and lost-time incident rates were lower at Amazon Robotics sites than at non-robotics sites ^[2].

These are company claims. Amazon has not published the number of sites compared, the rates themselves or how sites were matched, so the claim cannot be checked from public information.

The ILR Review study: fewer severe, more non-severe injuries

The strongest independent evidence comes from Gordon Burtch, Brad Greenwood and Kiron Ravindran, published in the ILR Review in August 2025 ^[3]. Studying Amazon's warehouses, the authors found that robotics were associated with a 40% decrease in severe injuries and a 77% increase in non-severe injuries ^[3]. They report evidence that the rise in non-severe injuries is at least partly due to the faster pace of work at robotic facilities ^[3].

George Mason University's summary of the study adds detail. The increase in non-severe injuries was sharpest during high-demand periods such as Prime Day and the winter holidays ^[4]. The researchers also analyzed thousands of online posts by Amazon warehouse workers. Workers at robotic sites described the job as less physically exhausting, but reported much higher performance targets, in some cases a pick rate two to three times higher than at non-automated sites ^[4]. The summary describes non-severe injuries as things like repetitive motion problems that often lead to reassignment or light duty rather than missed work ^[4].

Reveal's reporting: why the numbers differ

In September 2020 Reveal, from the Center for Investigative Reporting, published an investigation based on Amazon's internal safety records ^[5]. It found that at Amazon's most common type of warehouse, the sortable centers that ship small and medium items, the rate of serious injuries from 2016 to 2019 was more than 50% higher at sites with robots than at sites without ^[5]. Reveal also reported that injury rates spiked in the weeks of Prime Day and Cyber Monday ^[5].

The two findings are less contradictory than they look. Reveal defined serious injuries as work-related injuries and illnesses that require days away from work or a job restriction ^[5]. The ILR Review study separates severe from non-severe injuries, and its increase is in the non-severe group, which includes injuries that lead to restricted or light duty ^[3][4]. A definition that counts job restrictions as serious will pick up much of that increase. Both sources point the same way: robotics changed the type of injury, and pace of work was part of the story.

Evidence beyond Amazon

Two other studies are useful context, though neither is about warehouses alone.

• Manufacturing in the US and Germany. A 2022 NBER working paper by Gihleb, Giuntella, Stella and Wang found that a one standard deviation increase in local robot exposure in the US reduced annual work-related injury rates by about 1.2 cases per 100 workers ^[6]. US areas more exposed to robots also saw significant increases in drug- or alcohol-related deaths and mental health problems ^[6]. In Germany, a one standard deviation increase in robot exposure was linked to a 4% decline in physical job intensity ^[6].
• Cobots in South Korea. Jung and Yang (2025) found that South Korean manufacturing firms using collaborative robots had more industrial accidents. A higher share of long-tenured employees and job rotation reduced injury rates, while multifunctional training was associated with more injuries ^[7].

The common thread is that robots reduce exposure to some physical hazards, but the outcome depends on how the remaining human work is designed.

How robots injure people

Robots also bring their own injury patterns. A study of US robot-related deaths at work from 1992 to 2017 counted 41 deaths. Most involved stationary robots, 78% involved a robot striking the worker while operating under its own power, and many of those happened during maintenance ^[8].

For non-fatal injuries, an analysis of OSHA severe injury reports from 2015 to 2022 found 77 robot-related accidents. The 54 involving stationary robots mainly caused finger amputations and fractures to the head and torso. The 23 involving mobile robots mainly caused leg and foot fractures ^[9]. The authors call for guards and collision avoidance systems that detect individual body parts, not only whole people ^[9]. OSHA notes that many robot accidents happen during non-routine work such as programming, maintenance, testing, setup or adjustment, when a worker may be inside the robot's working area ^[10].

Analysis: for mobile robots, lower-leg and foot injuries point to hazards at floor level, such as being run over, pinned against racking, or stepping into a robot's path. Check that the robot's detection covers feet and low obstacles, not only a standing person.

What the evidence means for buyers

Scale makes these questions practical. Robotics and Automation News reports that DHL Supply Chain and Locus Robotics passed one billion picks with thousands of robots across more than 40 DHL-managed sites, with plans for 5,000 more robots ^[11]. At that scale, a change in pick rate affects thousands of workers. See our warehousing and logistics chapter and the Locus Robotics and Vecna Robotics profiles for more on warehouse automation vendors.

How to measure outcomes in your own pilot

Use the standards and your existing safety technology to get data the vendor will not give you by default. For mobile robots, ANSI/A3 R15.08 Part 3, published in September 2026, sets out user duties including risk assessment, management of change, training, and periodic reviews and inspections ^[12]. Under Part 2, a user who modifies a mobile robot or system takes on the role of manufacturer or integrator for that change ^[13].

1. Set a baseline. Record injury rates by severity, restricted-duty cases and musculoskeletal reports for the area for at least a year before the robots arrive, or pick a similar comparison site.
2. Track pace. Record pick rates and targets per person per hour, before and after. The ILR Review and Reveal findings both point to pace as the risk ^[3][5].
3. Split by severity. Report severe and non-severe injuries separately, plus restricted-duty days. A single recordable rate can hide the shift the studies found.
4. Use the robot data. Fleet managers log protective stops and interventions. Ask for export rights and treat stops near people as near misses.
5. Add an independent view. Some video analytics tools can check robot zones: Protex AI, for example, lists AGVs and AMRs among the vehicle types it recognizes ^[14]. Proximity systems are adapting too. ELOKON says its UWB AGV module has a detection range of up to 49 feet for avoiding collisions between AGVs and other trucks fitted with its system ^[15]. See our measuring outcomes chapter for how to set leading indicators.
6. Plan for vendor exit. When Zebra Technologies wound down its Fetch mobile robot group in December 2025, The Robot Report said the future of robots at customer sites likely depended on whether Zebra could sell the group ^[16]. Ask who will service the fleet if the vendor leaves the market.

What to do next

• Ask any robot vendor that claims injury reductions for the data, the sites compared and the injury definitions used.
• Add pick rate, restricted-duty cases and musculoskeletal reports to your pilot scorecard alongside recordable injuries.
• Test robot detection with a foot and a low obstacle, not only a standing person, during acceptance.
• Write export of protective stops and near misses into the purchase contract.
• Assign owners for R15.08 Part 3 user duties, especially management of change after layout moves.