Here’s a list that surprises people.
In Anthropic’s report What work can robots do? , nine of the ten most exposed occupations are vehicle operators. Taxi drivers top the whole index at 2.2 out of 3. Driving is mostly one task, done for hours, and autonomous vehicles can already do it on real roads.
Now the other end. Nursing and general repair are among the least exposed jobs. The report says present-day robots “can do little of their work even in highly controlled environments.”
Driving a two-ton car through a city is easier for a robot than giving a patient a shot or swapping a part under a sink. That tells you where the hard part is.
It’s the hands
Anthropic asked a simple question about every physical task: if robots could do it, what would still stop them? Then it sorted the answers.
- Capabilities block adoption for about 70% of physical tasks.
- Within that, manipulation stands out. Half of physical tasks wouldn’t be automated at scale unless robots get better at “touching and handling objects.”
- Planning and reasoning, the part AI models are most likely to improve, limits only about 8% of physical tasks.
Read those last two together. The thing everyone is pouring money into, smarter models, isn’t the main bottleneck. The bottleneck is a hand that can feel what it’s holding.
The examples are all about touch
The report’s examples read like a list of things that are easy for people and miserable for machines.
Dig trenches. A robot can cut a straight trench in open ground. Anthropic cites a system that retrofits excavators to do it on their own. But a big share of trenching is careful digging by hand around buried pipes and cables. Robots can’t do that, so overall the task is rated unexposed.
Finish drywall. A drywall robot can scan walls, spray additional coats and sand them. But pressing paper tape into wet compound over a joint? Mud and tape are messy. Robots can’t. The same job is part highly exposed and part untouchable.
Bleach, dye or tint hair. Fine motor control on a person’s head. Unexposed.
Untangle wires. Anthropic uses this as its example of the dexterity robots are missing.
Every one of those is a sensing problem as much as a thinking problem.
What touch really means in hardware
I spent the first part of my career in test and measurement and IoT. Sensing was the job. Here’s what I’d point out to anyone who thinks a smarter model fixes this.
A camera isn’t a fingertip. Vision tells you where the glass is. It doesn’t tell you how hard you’re squeezing it, whether it’s slipping or whether it’s wet. Anthropic’s footnotes point to exactly this split: some robots use cameras only, while others use touch sensors. Amazon calls its Vulcan robot its “first robot with a sense of touch” , and the report cites robots like it doing warehouse picking and stowing.
Sensors need bandwidth and a fast loop. Holding something fragile is a control loop: sense force, adjust grip, sense again. The more points you sense and the faster you close the loop, the better you do. For scale, Pollen Robotics’ $399 Microduck , a 25 cm open source biped, runs its onboard policy at 50 Hz and picks things up with an articulated beak. It walks, falls, gets back up and learns new tricks. It doesn’t have hands, and nobody would expect it to fold your laundry.
Hands have to survive the job. A hand with lots of fingers and sensors has lots of parts that can break. For now, people fill the gap. In MarketWatch’s test of two household humanoids, the robots vacuumed floors and wiped counters “with some help from humans behind the scenes” , with teleoperators at the company’s office steering every move. A person was still in the loop.
Who’s exposed, stated plainly
This is a heated topic, so I’ll stick to what the report says.
Workers in the most robot-exposed jobs are more likely to be male, less likely to hold a bachelor’s degree, and paid around $30 less per hour than unexposed workers. Their jobs are more likely to involve carrying weight, extreme heat or hazardous contaminants. Anthropic notes this is in many ways the opposite of who’s exposed to LLMs.
The report also backtests its method across 50 years. Jobs that were more exposed to the robots of their day saw wage and employment declines in later decades. That’s a real signal and I don’t want to wave it away.
But exposure isn’t the same as replacement. Remember the cost column : robots are cost-competitive on 0.3% of tasks today. And the most exposed jobs are exposed because of one task, like driving, not because a robot can do everything the person does.
What I’d watch
If you want to know when robots move into new kinds of work, don’t watch model benchmarks. Watch hands.
Watch for grippers that can tell slip from squeeze. Watch for tactile sensors that ship in products, not just in papers. Watch for hands that are cheap to repair. And watch the teleoperators. Every time a company can take a remote human out of a task, that’s a capability that just became real.
The brain is getting better fast. The hands are the part to watch.