The hardest thing for a robot to do is become useful

The Straits Times, The hardest thing for a robot to do is become useful

By Professor Mohan Rajesh Elara, Academic Director (Innovation PG), Office of Provost

The real test for a robot is not whether it can walk, dance or impress investors. It is whether it turns up for work every day, reliably and economically.

Two things happened in 2026. Unitree Robotics priced its Shanghai initial public offering (IPO) at a valuation of about 61 billion yuan (S$11.5 billion), putting it on track to become China’s first publicly listed humanoid robot manufacturer. Investor demand was extraordinary, with the retail tranche more than 8,000 times oversubscribed. Its shares ended their first day of trading on Aug 19 up by more than 460 per cent.

More quietly, Singapore became a super-aged society in 2026. More than one in five of us are now 65 or older. By 2030, it will be one in four.

Few countries have more riding on a single question: When does an impressive robot become a useful one?

I should declare an interest. I co-founded LionsBot, a Singapore company building autonomous cleaning robots, so I have watched this from the inside – and what follows is, in part, an argument for my own industry.

Unitree’s IPO is significant not because it proves that the humanoid problem has been solved, but because it shows that robotics is entering a different phase. Capital markets are beginning to fund the expensive transition from impressive prototypes to manufacturing, deployment and continuous learning at scale. The extraordinary investor demand reflects confidence in that future – but also expectations running ahead of present-day capability. An industry can be mature enough for public markets while its defining product is still unripe for everyday use.

My position, stated plainly, is that a general-purpose humanoid – one you could buy, switch on and turn loose anywhere – is at least a decade away. Robots will arrive all around us regardless, in almost every shape but that one. And Singapore’s advantage will come from two things at once: building robots that are genuinely useful, and becoming the first nation designed to work seamlessly with them – across our buildings, our precincts, our industrial parks and, in time, the city itself.

Why the last mile is so hard

For a digital artificial intelligence system, the world is made of information. For a robot, the world is made of physics.

A robot must contend with friction, weight, deformable objects, wet floors and things never quite where they were expected. Then we add people, who walk unpredictably, leave bags in corridors and change a process without telling the machine.

So, it must manage three shifting domains at once – the objects it handles, the space it moves through, the people around it – then work at close to 100 per cent reliability, at a price that makes commercial sense.

This is why the final 10 per cent of robotics development is harder than the first 90. A demonstration absorbs failure. A shift cannot. And it is why so much of this industry stalls between the machine that dazzles in a laboratory, the machine deployed on pilot funding, and the rare machine that simply does everyday work at scale. Only the third kind changes anything.

The future will not all be humanoid

Humanoids have captured the imagination, and their long-term case is real: doors, stairs and workstations all assume human dimensions and dexterity. But for most tasks, the best robot will remain the robot designed for the job. In many industries, the equipment people already use will simply become robotic: the manual scrubber becomes an autonomous cleaning robot, the material trolley an autonomous mobile one. It makes little economic sense to build a humanoid to sit in the driver’s seat of a self-driving car, or to push a machine that can already drive itself.

Physical capability is also not economic usefulness. Humanoids can run and dance; reliable manipulation in an unpredictable space is far harder. Over five years, I expect them in factories, where people already work in structured space.

Homes are different. A factory holds thousands of repetitions of one process; a home holds thousands of processes, each performed occasionally. Preparing breakfast, finding misplaced spectacles, helping an elderly parent, moving safely around children – that demands perception, dexterity, judgment and safety we cannot yet build. The Singapore household of the 2030s will hold many robots long before it holds one humanoid.

Who will do the punishing work?

Automation makes people uneasy, and that deserves an honest answer. Our situation, though, is unusual. We are not facing machines that displace a plentiful workforce, but the opposite: work that must still be done, and fewer people each year willing or able to do it.

Walk through a mall after closing and look at who is pushing the scrubber. Ride a service lift in a hospital and look at who is moving the linen. Many are older workers, some in their 60s and 70s, doing work that ages the body faster than the calendar does.

That is the honest case for robotics here: not to remove people from work, but to remove the work that wears people out. It is not automatic. It requires training, redesigned roles, and employers willing to pay for the better job that results.

The race is for the learning loop

Deploying a robot has traditionally demanded heavy engineering: mapping a site, writing rules, handling exception after exception. AI is compressing that cycle – fewer examples needed, unfamiliar spaces read more capably, millions of situations rehearsed before the machine meets the physical world.

But simulation takes you only to the door. Every robot genuinely at work generates what no laboratory can synthesise: a failed grasp on a bag heavier than it looked, a corridor blocked by a delivery trolley, a floor made treacherous by spilt bubble tea. That experience trains the next model, which improves the next fleet, which produces better experience still.

So, the race ahead is not to build the most capable machine. It is to close that loop fastest between robots, AI, data and the physical world – and a loop, once spinning, compounds. Whoever is a year ahead in deployment is more than a year ahead in learning.

A country designed to work with robots

Useful robots are only half the answer. The other half is the world we ask them to work in, and here we can do something no other country can.

Today, a robot arriving somewhere new must work everything out alone – doors, thresholds, access control, crowds, weather – carrying the whole burden of understanding. Imagine the reverse: places that meet the machine halfway and tell it what it needs to know.

Start at building scale, with a map handed over on arrival and doors that announce their state. Widen to the precinct, where a robot leaves one block, crosses a covered walkway and enters another without losing its bearings – a hospital campus, a university, a housing estate. Then the industrial park, where autonomous machines move between plants, warehouses and loading bays under one set of rules. Then the port, the airport, the transit network. And eventually a city in which the seams between those places stop behaving like walls.

That is not science fiction but a planning decision – precisely the kind this country has made before. We built a port ahead of the trade, a water system ahead of the drought, an airport ahead of the passengers and a fibre network ahead of the applications. Lay the ground before the thing that runs on it, and the thing arrives sooner and works better.

What robotics needs is modest by comparison: not a grand programme, but a common set of rules for how machines and places speak to each other – a shared way to request a map, to pass a controlled door, to hand a task from one fleet to another. Set it once, apply it first wherever the state builds, and let it spread.

Some of that groundwork is under way. At the Singapore University of Technology and Design (SUTD), my colleagues and I have been developing a Robot-Friendly Building Level of Service – a way to rate how ready a building is for robots, much as we already rate buildings for accessibility, energy and fire safety. It gives an owner a grade, a benchmark against comparable buildings, and a specific list of what to fix: a lift the robot can call, a door it can pass, a route that does not end at a kerb.

Then, the two halves compound. Places designed for robots make every deployment faster, which produces more data, which makes the robots better.

The jobs on the other side

There is a second thing we must build ahead of demand, and it is not made of concrete. Every wave of automation has ended the same way: the work of one decade becomes the work of the next, and the people who do well are those who learnt to command the new tools rather than compete with them. AI is doing this faster, and to more occupations at once, than anything before it. For a small country, the question is not whether it happens, but whether our graduates arrive on the right side of it.

That is why this work has to be anchored in real sectors rather than in the laboratory. With support from the National Robotics Programme, the Civil Aviation Authority of Singapore, and the Building and Construction Authority, we at SUTD are developing robots to address critical challenges across two key sectors: built environment and aviation. Both are labour-short, physically punishing and stubbornly hard to automate – which is precisely why they are the right places to work. The requirements come from the enterprises, contractors and operators who will run the machines, not from a research agenda. And the students who build them graduate having solved a problem an industry actually has, in a sector the economy actually needs.

That, in the end, is what my colleagues and I now spend our time teaching. Not robotics as a subject, but how to work alongside these machines and turn them to account: real domain knowledge in a real sector, the AI and engineering skill to build, and the judgment to know which problems are worth automating at all. Graduates with that combination will not be displaced by this technology. They are the ones who will decide how it gets used.

Measure usefulness, not spectacle

We should resist judging progress by demonstration. The better questions are duller. How many hours does the machine work without intervention? Can an ordinary worker operate it? And hardest of all – does anyone still want it once the novelty has gone? Judge the country the same way: not by the pilots we launch, but by how many graduate into permanent work.

The listings will supply the capital, AI the intelligence, better hardware the body. None of them can supply the ground it all must stand on. Build that first, and we will export more than robots. We will export the blueprint for a country that works with them.

The robot revolution will not arrive on the day a humanoid walks convincingly across a stage. It will have arrived when robots are all around us, doing useful work – and we no longer notice them at all.

• Mohan Rajesh Elara is a chair professor and academic director (innovation postgraduate) at the Singapore University of Technology and Design, where his research focuses on robotics and physical AI.