5 March 2026 · LinkedIn
In 2022, five divers were pulled into an oil pipeline off the coast of Trinidad and Tobago by a pressure differential during routine maintenance. Four of them died. No rescue attempt was made - the same hazard that killed them made any intervention too dangerous. A subsequent inquiry found grounds to pursue the pipeline's owner for gross negligence and corporate manslaughter.
There is no version of that story that ends well. Just four people who went to work and didn't come home, in an industry that knew the risks and sent them anyway.
I've thought about it a lot while watching what's been quietly happening in robotics.
For decades, building a robot that could function in genuinely unstructured environments - not a controlled factory floor - was the hard problem. Every eventuality had to be pre-programmed. Change the conditions and the robot was lost.
What changed is the same architecture behind the AI tools everyone started using in 2023. AI applied not to language, but to movement. The ability to read an environment dynamically and respond to the unexpected. Predicting what movement makes the most sense next.
Figure AI's humanoid robots are running 10-hour shifts on BMW's production line in South Carolina. Figure 02 picks up metal sheets, walks them across the plant, places them with millimetre precision into welding fixtures – seeing, deciding, acting, without a human guiding its hands. Their Figure 03 is even more advanced. Boston Dynamics' Atlas is on Hyundai's line. And there's a host of competitors, still. Goldman Sachs puts the humanoid robot market at $38 billion by 2035.
The race to deploy capable humanoid robots at commercial scale is not a future story.
Three years of AI and jobs conversation has focused almost entirely on knowledge work. The analyst, the programmer, the creative. That disruption is real, and it’s largely a conversation knowledge workers have been having about themselves. Manual labour barely features. Yet manual labour employs, globally, a vastly larger number of people.
Some of that work is dangerous. Pipelines inspected by people who shouldn’t have to be inside them. Mines worked by crews who accept the risk as the price of the job. Truly unstructured environments – underwater, confined, completely unpredictable – remain the hardest problems in robotics to crack. But the direction of travel is unmistakable, and it can’t come quickly enough.
The exposure isn’t uniform. Trades navigate a different environment on every job: diagnosis-heavy, spatially unpredictable, built on practical judgement. They have a resilience that factory work doesn’t. Assembly lines, warehouse operations, logistics – structured, repetitive, scalable. That’s where pressure lands first. It accounts for a very large number of people.
The case for robots in dangerous work is straightforward. What happens to the much larger workforce doing work that doesn't kill them, just employs them - that question hasn't seriously been asked yet.
Join the conversation on LinkedIn →
← All writing
