The Robotics Bottleneck Is Moving Up the Stack

One presentation argued that robot learning needs to move from copying movements to understanding purpose.

Another showed that the robotics industry is still dominated by conventional industrial machines, with millions already deployed and factories demanding reliability, uptime and ROI.

These are not competing narratives.

They describe two different layers of the same market.

And the investment opportunity may sit in the layer connecting them.

A Munich-based physical-AI researcher presented a progression in robot learning, from direct mimicry toward what he calls purposive learning.

The distinction is important.

Most imitation-based systems learn from demonstrations: observe a human, collect trajectories, and train a policy to reproduce the observed behaviour. That approach can work extremely well inside a defined task distribution.

But it creates a scaling problem.

Consider three people cutting bread. One uses her right hand. Another uses her left. A third rotates the loaf while cutting.

At the trajectory level, these are different demonstrations.

At the task level, they are the same: cut the bread.

The proposed abstraction is therefore not "what exact motion did the human make?" It is: what was the human trying to accomplish?

The research agenda behind this idea goes beyond imitation, toward representations involving objects, tools, actions, goals and affordances. The objective is a compact representation of the task that can then be translated into an embodiment-specific policy.

This could matter economically. If robot learning moves from trajectory memorisation toward task-level representation, three things potentially change.

Data efficiency. More demonstrations can contribute to the same underlying task representation, rather than simply expanding a library of motion trajectories.

Embodiment transfer. The same task representation could be adapted to different arms, hands, mobile manipulators or humanoids.

Deployment time. Engineering effort could shift from manually programming every motion and exception toward specifying the task, its constraints and its success conditions.

The demonstrations presented at the conference were striking. The researcher described a dexterous ball-swapping task that he estimated would require roughly 1.5 years of conventional engineering work, against approximately 3.5 weeks of live learning. He also described a humanoid policy trained overnight for a Beijing half marathon, and a demonstration in which a newly delivered robot was cutting bread by the afternoon.

These should not yet be read as production benchmarks.

The important question is not whether a robot can learn one impressive demonstration quickly. It is whether the same approach survives different objects, different lighting, different tooling, different workers, contact uncertainty, safety constraints, hardware variation, long-duration operation and failure recovery.

That is where the research becomes commercially interesting.

The real benchmark is not time-to-demo. It is time-to-reliable-deployment.

The second presentation came from Alexander Verl, Chair of the Research Committee at the International Federation of Robotics.

The numbers are a reality check for the humanoid narrative.

The IFR recorded 542,076 industrial robot installations globally in 2024, with more than 4.66 million industrial robots already in operation — annual installations more than double the level of a decade earlier.

China is the centre of gravity. It installed approximately 295,000 industrial robots in 2024, representing 54% of global installations, and its operational stock exceeded two million units, the largest installed base in the world.

The market is also still heavily concentrated in conventional industrial robotics. Articulated arms, SCARA, Cartesian and Delta robots remain the dominant platforms.

And the largest application category is not "general-purpose intelligence."

It is handling.

The implication is straightforward. Physical AI does not start with humanoids. It starts with an enormous installed base of machines that already perform useful work.

Humanoids are trying to expand the addressable task space. AI-enabled industrial robotics is trying to make the existing installed base substantially more flexible.

Those are different opportunities.

Robot density makes the divergence clearer.

South Korea leads the world at approximately 1,220 industrial robots per 10,000 manufacturing employees. China sits at 166, despite operating the world's largest robot stock.

The denominator matters here — China has an enormous manufacturing workforce, and density rewards small industrial bases. But the broader point stands: there is still significant room for conventional automation before humanoids need to replace the industrial arm.

And adoption itself is slow.

Collaborative robots are a useful precedent. They have grown into a meaningful portion of the market, but it took roughly 20 years to reach around 12–13% of installations, according to the IFR presentation at WRC.

This is why robotics is fundamentally different from software.

A model can be upgraded overnight. A factory cannot.

Deploying a new robot often means integrating hardware, fixtures, perception, controls, safety, factory software, operators and maintenance. The cost of the robot itself can therefore be only one part of the deployment equation.

For an industrial customer, the relevant question is not: can the robot perform the task?

It is: can the robot perform the task reliably enough to justify changing the production process?

This is where the two presentations become particularly interesting.

Industrial connectivity is not the missing piece. Standards such as OPC UA already allow machines and software systems to exchange structured information.

The more interesting missing layer is the representation of the task itself.

Today, much of industrial robotics still looks closer to: integrator → robot-specific programming → fixed workflow.

The potential transition is toward: task specification → learned policy → deployment across compatible embodiments.

That distinction is enormous.

If a factory teaches a robot how to perform a task once, and that task representation can subsequently be adapted to another robot, another production line or another SKU, the economic value of the software layer rises dramatically.

The robot becomes the execution substrate. The task representation becomes the reusable asset.

This is where I would be most cautious about the current Physical AI narrative.

A robot that learns to cut bread in an afternoon is impressive. But the commercial question is: how many edge cases does it encounter before production stops?

A factory robot has to deal with objects outside the training distribution, partial occlusion, changing tolerances, worn tools, human intervention, unexpected collisions, sensor degradation, production-line variation, and recovery after failure.

The final 5% of reliability may be substantially harder than the first 95% of capability.

This is analogous to autonomous driving. Getting a system to perform a task in normal conditions is one problem. Building the perception, planning, validation, monitoring and recovery infrastructure necessary for deployment at scale is another.

That is why the most valuable Physical AI companies may not be the ones demonstrating the most spectacular behaviours. They may be the ones solving the boring problems: validation, deployment, observability, recovery, interoperability and fleet learning.

The robotics industry has spent decades optimising cost per robot.

AI may shift the industry toward optimising cost per deployed task.

That is a much more interesting metric.

Imagine two robotic systems. System A costs $50,000 and requires four months of engineering integration before entering production. System B costs $80,000 but can learn, validate and deploy a new task in two weeks. System B may be dramatically cheaper on a total-cost-of-ownership basis.

This suggests a different set of metrics for evaluating Physical AI companies:

These tell us much more about commercial scalability than a humanoid robot completing another impressive demo.

There are currently two clocks running in robotics.

The AI clock is moving quickly. Models are becoming better at perception, reasoning, planning and learning from demonstrations.

The industrial clock is moving much more slowly. Factories care about reliability, safety, uptime, integration cost and ROI.

The investment opportunity is likely somewhere between them.

The humanoid is the most visible manifestation of Physical AI. But the more interesting platform question may be: what software and infrastructure allows intelligence to move from a model into a physical task — reliably, across heterogeneous machines, at economically viable cost?

If that layer works, humanoids become one embodiment of the system rather than the system itself. And the addressable market suddenly includes the millions of industrial robots already deployed, not just the next generation of humanoids.

That is the distinction I would watch.

The future of robotics may not be about building a robot that can do everything. It may be about building the infrastructure that lets any suitable robot learn to do more things, faster, with less engineering.

That is a much larger platform opportunity.

Watch the full World Robot Conference stream: https://www.youtube.com/watch?v=r39i7_LTPOA

#PhysicalAI #Robotics #Humanoids #AIInfrastructure #WorldRobotConference

Carol Chen
Founder, Compute Notes
Builder of AI-native businesses and investor in AI infrastructure

LinkedIn: linkedin.com/in/carol-c-76461498
Website: computenotes.co
Network: network.computenotes.co

Get a free AI spend teardown → watt.computenotes.co