whirrwork
Insights

How robot jobs go wrong: a field guide to real failures, and what each one teaches buyers

Drone shows that fell, robotaxis that misjudged, a security robot in a fountain, a fatal commissioning accident. Real, sourced incidents, the pattern behind them, and what buyers should ask before signing.

By The Whirrwork team9 min read
  • robot safety
  • incidents
  • drone show accident
  • robotaxi
  • warehouse safety
  • risk management
  • due diligence

The short version

  • Most robot failures we studied weren't robots "deciding" anything. They were setup errors, environments outside the spec, or edge cases nobody planned for.
  • The most dangerous moments are often commissioning, testing and the first minutes of an operation, when people are close and checks get skipped.
  • The best operators are the ones with written checklists, a second person who verifies, and an honest incident history. Ask for all three.
  • Every incident below is real and linked to an investigation report or reputable coverage. Where people were hurt, we've kept the tone plain.

We're pro-robot. That's exactly why we read accident reports. A robot industry that hides its failures can't fix them, and a buyer who doesn't know how things fail can't ask the right questions. So here's the field guide.

1. Drone shows: the configuration error

What happened. On 21 December 2024, during a 500-drone holiday show at Lake Eola in Orlando, several drones collided and fell into the crowd. A seven-year-old boy was seriously injured (AP).

Why, per the NTSB. The final report found the uploaded show centre was off by about 5 metres and the show heading was rotated by about 7 degrees. When the show centre was updated, only 498 of the 500 drones were selected. A change in the company's geofence margin policy hadn't been applied to the laptops in use. The NTSB noted the operator "lacked a formal requirement for independent verification of critical show parameters before launch" (NTSB final report DCA25LA065).

The lesson for buyers. The drones didn't fail. The process did. Ask your show operator: who independently checks the show centre, heading, geofence and drone count before launch, and do they sign anything? Ask to see the crowd setback distance on a map.

2. Drone shows: the weather nobody was watching

What happened. On 14 July 2023 in Melbourne's Docklands, a 500-drone show before a Matildas match went wrong shortly after launch. 427 drones ended up in the water; divers recovered 236. No one was injured (ATSB investigation AO-2023-033).

Why, per the ATSB. The drones met winds beyond their published capability, and later more than twice the limit. The pilot didn't know the ground station displayed wind speed, and the software didn't actively alert on exceedances. The operator has since required two CASA-approved pilots per show and single-drone wind test flights beforehand (ATSB, July 2025).

The lesson. Ask for the wind limit of the specific drone model, who decides go/no-go, and whether there's a test flight on the night. A good operator will cancel. Put the weather-cancellation terms in the contract so they're not tempted to fly anyway.

3. Drone shows: someone else's interference

In October 2018, a 100-drone show over Hong Kong's Victoria Harbour went wrong when GPS signals were jammed; 46 drones fell into the harbour, with about HK$1 million in damage, and police opened a criminal investigation (South China Morning Post).

The lesson. Robots that depend on GPS or radio links can be knocked out by things outside the operator's control. Ask what happens on signal loss (hover, land, return?) and whether the launch area is over water, people or empty ground. Over water, a failsafe landing costs drones. Over people, it costs much more.

4. Robotaxis: what happens after the first collision

What happened. On 2 October 2023 in San Francisco, a pedestrian was struck by a human-driven car and thrown into the path of a Cruise driverless vehicle. The Cruise car braked but hit her, then attempted to pull over and dragged her about 20 feet (NHTSA consent order; Cruise recall filing).

What followed. California's DMV suspended Cruise's driverless permits, citing both vehicle safety and misrepresentation of information (California DMV). NHTSA later fined Cruise $1.5 million for incomplete crash reports.

The lesson. Two of them. First, the "after an incident" behaviour of a robot matters as much as avoiding the incident. Second, disclosure is a safety feature. A supplier that minimises a problem to a regulator will minimise it to you. Ask for their incident log and reporting process in writing.

5. Robotaxis: the software update that went backwards

In December 2025 Waymo recalled 3,067 vehicles because a software release could let a stopped car proceed past a school bus with flashing lights or a stop arm. The issue was in software from an August 2025 release; Waymo said all vehicles were updated by 17 November 2025, and no collisions were reported (NHTSA recall report 25E-084).

The lesson. Robots change every time their software does. For any fleet you rent, ask: how are updates tested, can I see release notes, and can I delay an update during a critical period? A robot that was safe in March isn't automatically the same robot in September.

6. Drone delivery: the crane that wasn't there at breakfast

On 1 October 2025, two Amazon MK30 delivery drones flew into the boom of a mobile crane in Tolleson, Arizona, about nine minutes apart. No one was injured. The crane had been erected between 7:30 and 8:00 a.m.; Amazon's ground crews had done their rooftop obstacle scan at 6:58 a.m. (NTSB preliminary report WPR26LA001). Amazon said it added "enhanced visual landscape inspections" for moving obstructions (NBC News).

The lesson. Site surveys go stale. If you're hiring drone inspection, mapping or delivery near a construction site, tell the operator about cranes, scaffolds and temporary structures, and ask how often they re-check the route.

7. Security robots: the floor matters

In July 2017, a Knightscope K5 security robot patrolling a Washington, D.C. office complex rolled down steps into a fountain. Knightscope's investigation said its wheel-slip algorithm failed to detect a skid on a "loose brick surface" (Security.World; Reuters). Nobody was hurt, and the internet had a lovely week.

Here's the part buyers should note: under its subscription model, Knightscope said a replacement robot was on site within 48 hours and the client was credited a month's subscription. That's what a good service remedy looks like.

The lesson. Walk the route with the supplier. Steps, water, ramps, loose surfaces, glass doors. Then get the replacement time and service credit in the contract, because something will go swimming eventually.

8. Delivery robots: blocking the people the sidewalk is for

In October 2019, a Starship delivery robot at the University of Pittsburgh sat in a curb ramp and blocked a wheelchair user, Emily Ackerman, from getting onto the sidewalk. The university paused the programme, and Starship updated its mapping of the intersection (The Pitt News; WESA).

The lesson. "Works fine" usually means "works fine for the people who tested it". If a robot shares space with the public, ask how it handles wheelchairs, strollers, guide dogs, crowds and kids, and who people can complain to.

9. Industrial robots: commissioning is the dangerous phase

In November 2023, a worker in his 40s was killed at a vegetable packing plant in South Korea while checking the sensors of a robot that moved boxes onto pallets; the robot gripped and pressed him against a conveyor (AP).

A 2024 study of US OSHA Severe Injury Reports found 77 robot-related accidents from January 2015 to April 2022. Stationary robots caused 66 injuries, mostly finger amputations; mobile robots caused 27, mostly leg and foot fractures (Applied Ergonomics, via PubMed).

The lesson. The robot is most dangerous when someone is inside its working area: testing, clearing a jam, doing maintenance. Ask your integrator how lockout, safety-rated stops and speed limits work during commissioning, and which standard the cell is built to. In 2025 the core industrial standard, ISO 10218, was revised, and collaborative-robot requirements were folded into Part 2 (ISO 10218-2:2025).

10. Warehouses: the robot changed the pace, not just the task

In 2020, Reveal from the Center for Investigative Reporting analysed Amazon's internal injury records and reported that serious injury rates at its robotic warehouses were more than 50% higher than at non-robotic ones from 2016 to 2019. The reporting linked this largely to faster pick rates and more repetitive motion (The Verge, summarising Reveal).

The lesson. A robot can make the human job harder. If you automate the walking, the people who remain may do more lifts per hour. Measure strain and injury rates before and after, not just throughput. More on this in robots and jobs.

11. The vendor disappears

In December 2024, Embodied, the maker of the Moxie companion robot for children, shut down. Because the robot depended on the company's cloud, owners were told it would stop working, with no refunds (Ars Technica).

The lesson. Ask what still works if the supplier goes bust or switches off its servers. For business contracts, ask about source-code or software escrow, data export and an offline mode. We go deeper in RaaS contract red flags.

12. Humanoids: the demo and the marathon

In April 2025, Beijing held a half marathon with humanoid robots. The winner, Tiangong Ultra, finished in 2 hours 40 minutes, fell once and needed three battery changes. Only six of 21 robots finished (Wired). Nobody was hurt, and honestly, most of us would also need three battery changes.

The lesson. Ask for runtime per charge, swap time and intervention rates in your conditions. Not the highlight reel.

The patterns

What went wrongIncidentsWhat to ask the supplier
Setup or configuration errorOrlandoWho independently verifies the setup before launch?
Environment beyond specDocklands, Knightscope, TollesonWhat are the limits (wind, slope, surface, obstacles), and who calls it off?
Interference or signal lossHong KongWhat does the robot do when it loses GPS or link?
Edge cases in public spaceCruise, Waymo, StarshipHow are rare situations tested, and how are updates rolled out?
People inside the working areaKorea, OSHA studyHow do lockout and safety-rated stops work, to which standard?
Human job got harderAmazon (Reveal)What happens to the pace and strain of the remaining jobs?
Vendor failureMoxieWhat still works if you go away?
DisclosureCruiseShow me your incident log and how you report to regulators

Bot or Not?

A robot is a safe bet when…Be careful, or hire humans, when…
The operator has written checklists and a second-person checkIt's "just one pilot, he's done this loads"
You can control the area (fences, setbacks, closed site)The public can wander into the working area
The supplier shares incident history openlyThe supplier says they've "never had an incident" (everyone has)
Weather and site limits are written downNobody can tell you the wind limit

What to ask, in one list

  1. Your incident history for the last three years, and what you changed after each.
  2. The written pre-operation checklist, and who signs it.
  3. Environmental limits for this exact model, and the cancellation policy.
  4. Failsafe behaviour on signal loss or fault.
  5. Safety standard(s) and certifications that apply here.
  6. Insurance certificate naming us (see liability and insurance).
  7. Replacement time and service credits if the robot fails.
  8. What still works if your company or cloud goes away.

Then see our full buyer's due diligence checklist.

Sources

All accessed October 2026.

More Insights