Any task. Any robot. Operator included.

Customer story

A robot in the stockroom.

Mary wanted help steaming garments. We brought a wheeled robot into her Singapore stockroom and operated it remotely.

Read Mary’s story

The same work turns up in every kind of business.

Fold, sort, pack, stock, count. Pick your setting and the examples change. The robot, the model and the stack stay the same.

How it works

Any task can go round this loop.

A person drives the robot first. Every session of real work teaches the model. Each lap, the robot needs them less.

01 / The job

Today

Start with your standard.

How it is held. Where it goes. When it is done right. The person who knows the work sets the bar.

02 / Human in the loop

Today

A remote operator drives. The robot records.

The operator connects through robot-stack from anywhere, even another country: about 8 ms from Malaysia to Singapore, about 60 ms from India. Every session of real work becomes data about this task.

03 / Post-train

In build

The model learns this job.

We post-train a model on that data, for this task, in this place. It has to meet your standard before it works alone.

Autonomous folding · demonstration · 10× speed

04 / Deploy and step in

In build

When it slips, a person takes over.

The robot works with an operator on call. Each intervention shows it the fix. Those moments are the most valuable data we collect.

Laundromat · commercial laundry · field work

05 / One operator, more robots

Target

Each lap, the robot needs them less.

As the robot handles more of the job, one operator can look after several. That is what makes robot labour affordable.

06 / The next task

Target

Good at one task. Next, the one beside it.

Skill in one task opens the ones beside it. Same robot, same place, a new lap of the loop.

FoldingSortingPackingStocking

More work.
Less hands-on time.

You pay for work done to your standard. As autonomy grows, each operator’s time goes further, and the cost of that work comes down.

Start with one task →

The deployment harness

Any model.
Any robot.
One harness.

robot-stack sits between AI models, robot hardware and the work on site. A coding model needs an agent harness to do real work. Robots need one too.

AI modelsswap the model

robot-stackWhere models meet real work

Run

  • Policy runtimeAny model, one interface
  • Device driversEach robot body plugs in

Guard

  • Safety checksLimits around every model
  • InterventionAn operator takes over remotely

Learn

  • RecordingEvery session and correction is kept
  • Post-trainingThe model learns this job

For your site

Different robots suit different work. We fit the robot to your job, and a person backs it up.

For builders

Bring your model or your robot. We put it to work on real jobs and show you how it holds up.

For the long run

Each better model ships into deployments that already run. What we learn on one task carries to the next.

The line

Build your
robot team.

Choose the robot, the model and how many. Or leave it to us: different robots suit different work.

Put a robot to work

Writing

Enterprise / Built around your operation

Start with one useful task.

We bring robots, software and human support to your operation. Together we test a task that matters, and build towards supervised autonomy.

Work in view

Four real tasks. Each one labelled for what it is.

For Singapore retail

Where would you start?

Most outlets have one job that keeps good people in the back room. Pick it.

Stockroom

  • Garment steaming
  • Fitting-room returns
  • Online order sorting

Back room

  • Boxed stock put-away
  • Stockroom picking
  • Returns sorting

Deliveries

  • Back-room unpacking
  • Crate and tote handling
  • Shelf-ready sorting

The work agreement

Agree the job. Earn the next step.

01

Define the job.

Agree the task, item mix and finish standard. Make handoffs and acceptable human help explicit.

02

Test it in your workflow.

Run the task in its real setting. Include loading, resets and recovery in the result.

03

Build towards autonomy.

Develop and evaluate policies against the same standard. Introduce supervised autonomy as requirements are met.

TRC brings

Robot hardware, our robot-stack software, integration and operating support.

You bring

Workflow knowledge, site access and agreed handoffs.

What counts as progress

Four numbers we agree before the robot arrives.

Accepted output

Quality and finish

Useful pace

In your workflow

Human effort

Every assist and reset

Full cost

Per accepted unit

Customer story / Singapore

It starts with Mary.

Mary wanted more of her team’s time on the shop floor. We brought a robot into her stockroom to steam garments, with an operator in control.

From human operation to supervised autonomy.

Today, people guide the work. We’re building towards a robot that handles more of Mary’s steaming, with human help when it matters.

01 / The job

Today

Start with Mary’s standard.

How the fabric is held. Where the steam goes. When the garment is ready. The person who knows the work sets the standard.

02 / Human expertise

Today

Human judgment. Robot movement.

People guide the early work. Through robot-stack, an operator brings dexterity and judgment to the robot.

03 / Useful work

Today

Useful work, in the real world.

A useful task, in the place it needs to happen. Garment steaming gives us a concrete job to improve.

04 / Task experience

In build

Turn experience into autonomy.

Demonstrations show how the work is done. We’re connecting them to training and evaluation, so new policies earn their place on the job.

01 / The job

Today

Start with Mary’s standard.

How the fabric is held. Where the steam goes. When the garment is ready. The person who knows the work sets the standard.

02 / Human expertise

Today

Human judgment. Robot movement.

People guide the early work. Through robot-stack, an operator brings dexterity and judgment to the robot.

03 / Useful work

Today

Useful work, in the real world.

A useful task, in the place it needs to happen. Garment steaming gives us a concrete job to improve.

04 / Task experience

In build

Turn experience into autonomy.

Demonstrations show how the work is done. We’re connecting them to training and evaluation, so new policies earn their place on the job.

ThenNext lap.

05 / Supervised autonomy

Target

More autonomy. The same standard.

  1. 01Robot worksA task policy handles the steaming.
  2. 02A person helpsAn operator corrects an exception.
  3. 03Robot resumesControl is deliberately handed back.

05 / Supervised autonomy

More autonomy. The same standard.

01

Robot works

A task policy handles the steaming.

02

A person helps

An operator corrects an exception.

03

Robot resumes

Control is deliberately handed back.

Writing

Notes from the floor.

Robots, the work they do, and what that could mean for the people running a business.

← All writing

11 November 2026

When the Models Are Really Good

What changes when a robot needs less of a person’s time?

The question I get asked the most is some version of this. Aren’t you a services business?

I get why it feels like that. Today we put a robot into a business and a person operates it from somewhere else. One person, one robot. Freeze it there and I agree with you. It looks like a staffing agency with expensive equipment.

But I think the question comes from today. It comes from not yet living in the world where the models are really good and the robots are really good. I’ve been living in that world for a while now. This is what it looks like from inside.

Two things are a given.

Start with the models.

In April last year Dyna let a pair of arms fold napkins for 24 hours straight and reported 99.4%. One task, one room, their own test.

This April Physical Intelligence put its new model on a pair of arms it had never folded a shirt with. It got 80%. Their expert operators, trying it on those arms for the first time, got 80.6.

In July Sunday Robotics reported 99.1% across 785 attempts. Nine kinds of garment, in homes the robot had never seen, with nothing tuned for the home.

These are company numbers. Nobody has audited them and you should read them that way. But look at the direction. Folding was the task everyone used to explain why robots are hard. In a little over a year it went from a research problem to a deployment problem.

Now the robots.

Unitree’s IPO filing says the average price of its humanoids went from about $85,000 in 2023 to about $25,000 in 2025. Goldman Sachs wrote in 2024 that the cost to build one had dropped 40% in a year, when their own analysts expected 15 to 20. The arm we deploy costs about US$2,000.

The line in that filing I keep going back to is a different one though. 73.6% of Unitree’s humanoid revenue came from research and education. The robots got cheap, and they mostly went to labs and classrooms.

So price isn’t what is keeping robots from work.

I’ll be honest, none of this is contrarian. Everyone I meet believes the models and the hardware are going to get really good. I think that is the contrarian approach now. To be not that contrarian. To take both as a given, look past the noise, and just focus on deployments.

99% is where the real work starts.

That’s the bare minimum, actually. At 99%, 1,000 items a day is still ten failures a day, and the customer needs all ten fixed. They’re not paying for 990. They’re paying for work done to their standard.

Models · Cloth folding
99.4%99.1%
Apr 2025Napkins,
one room
Jul 2026Nine garment types,
homes never seen
The number held while the job got wider. In between, a model folded shirts on arms it had never folded with: 80%, level with expert operators.
Folding has gone from a research problem to a deployment problem.
Dyna · Physical Intelligence · Sunday Robotics · Company-reported
Robots · Price
$85K$25K
2023Unitree humanoid,
average price
2025, Jan to Sep73.6% of revenue from
labs and classrooms
Industry-wide, the cost to build one fell 40% in a year. The arm we deploy costs about US$2K.
The robot is no longer the expensive part. The operator is.
Unitree prospectus · Goldman Sachs, 2024 · Our purchase price
Where we sit
99% is where the real work starts.
That last 1% is what we are built for.
1,000 items a day, at 99%
10 still fail. The customer needs all ten fixed.
Company-reported figures. Sources linked in the text.

In August Dyna published a number I think more people should look at. They took two generations of their model to customer sites that neither had seen. Graded against what the customer would accept, the older one passed 46% of the time. The newer one passed 87%. In their own lab both were at about 100.

About 100 in the lab. 87 at the customer. And 87 was a huge jump.

I think that gap is the business. The models will keep getting really good, and there will still be a last stretch at the customer’s site that a bigger model doesn’t close. It closes when somebody knows the job. Even something as small as where to grasp a stack. You grasp it in the middle, right? Grasp it by the corner and it’s fine in the demo, and then on a normal day it falls.

A lot of deployments fail right here, and I think for two reasons. Companies promise autonomy too early, without a stack that can handle the long tail. And that stack is not just software. The stack is real people. Or they pick the wrong task. Pick the wrong task and you can’t bring the interventions down, you never hit autonomy, and you fail in front of the customer even though you’re building in the right direction.

The agent builder.

So who does that work?

There are model makers and there are robot makers. Some companies are both. They call themselves research-driven deployment companies. Dyna is the best of them and I think they’re very, very good.

I think there’s a third thing to be, and we don’t have to look very far to see it. Look at language models. The model makers got really good, and a whole set of companies grew up on top of them building agents. In February Physical Intelligence described the same thing for robots. They called their models a layer that other companies build on. One partner in that post folds laundry for businesses, with remote specialists stepping in to fix a fold. With that partner’s data in the model, interventions dropped by half.

That is what we are. The agent builder. We don’t pre-train a model and we don’t manufacture a robot. A good agent has to be model agnostic, especially if it’s going to take a physical form, because the cost of upgrading it once it’s at work is extremely high. So robot-stack takes any model and any robot.

But you cannot just be a software provider. You have to deploy the robots. That’s the difference. That is the defining difference.

Two numbers.

Once the model is good and the robot is cheap, the expensive part is the person. If someone controls a robot for a whole shift, you’re still paying for a whole shift of their time, plus the robot.

So two numbers decide what kind of company this is. How much of a person’s time each robot needs. And how much engineering each new deployment needs. If those two don’t fall, the people asking me that question are right.

I think they fall. I don’t think it’s true that this doesn’t scale. It scales by bringing the operator remote, and then it scales by bringing the cost of remote down.

The robot has to be on site. The person helping it may not. We measure about 8 milliseconds from Malaysia to Singapore and about 60 from India. Today one operator drives one robot. Next the model learns the job from those sessions, and the operator stops driving and starts stepping in. Every time they step in, the robot is shown the fix. One to one becomes one to three. Then one to ten. People on this team have trained a hundred teleoperators and run the operations behind a robot foundation model before.

Waymo says it has about 70 remote assistance people on duty at any time, for a fleet of 3,000 cars. That’s one person for every forty-odd cars, and they’re not even driving. The car asks and they answer. Waymo didn’t start there either.

The second number falls a different way. Every robot we deploy runs on the same control stack, and the training we give an operator doesn’t change from robot to robot. What we learn on one task carries to the next. All of it is to bring the cost of the work to par with labour, and then below it. Even at par, the robot works much longer hours.

And the corrections stay with us. Real interventions at a real customer are the hardest data in this industry to get. You can’t buy them. You get them by winning the customer and then doing the job reliably every single day, which is a whole other ball game from putting cameras on people’s heads.

The last 1%.

I didn’t start out believing any of this. I came from a world that was trying to bridge the gap between simulation and the real world. And I was a sceptic that scaling teleoperation was the way to bridge the other one, the gap to real-world deployment. Then I went to the US to work on humanoids and saw what 1X was doing with a person behind the robot. I found it completely captivating. I also learned I’m not a lab person. What I love is the moment a technology is about to enter the market.

In Singapore we have a word for the work I’ve been describing. Saikang. The unglamorous jobs nobody wants and somebody has to do. Ten failures a day, fixed one at a time, written into an SOP, taught to an operator, fed back to a model. That’s how ten a day becomes one. That’s how 99 becomes 99.9.

When the models are really good, everyone gets the 99. The 99.9 you have to earn. We’ve already started.

One robot, one operator and a rack of garments in Mary’s stockroom in Singapore. Three robots to one operator is next.

Singapore boutique · teleoperated

Read Mary’s story →

Start here

What needs doing?

Tell us the work in your own words. Any task, any robot, any model. One stack carries it.

Where

Which movement is it closest to?

Drag or swipe

We’ll pick the robot and the model for you.

That is all we need. Or choose them yourself.

Now build it.

Illustration of your setup

02Which robot?

Any robot, or let us pick

03Which model?

04How far do you want to take it?

1 operator :1 robot · where every job starts today

Who do we talk to?

Building robots or models, or here as an investor or press? Write to us →

Your requestYour work
Movement
Volume
Robot
Model
How many
Trial
3 months

Got it. We’ll get back to you by the next working day.

  1. A person reads what you sent.
  2. We reply with a trial plan and a start date.
  3. The robot and its operator arrive and begin with your standard.
Work
Not described yet
Robot
TRC picks the best fit
Model
TRC picks the best fit
How many
1 robot

Try a robot for 3 months.

Describe the task, pick any robot, and we’ll get back to you by the next working day.