.jpg "Robotics") |
| The future of robotics may not look like us. |
The next great wave of automation won't look like science fiction. It will look like a very strange arm reaching places a human never could.
Every decade or so, the technology industry finds its new organizing metaphor. In the 1990s, it was the platform. In the 2000s, the network. In the 2010s, SaaS — software-as-a-service — became the template through which an entire generation of founders, investors, and operators understood how to build enduring businesses: sell subscriptions, expand within accounts, compound on recurring revenue, and build a moat from data and switching costs.
The next decade of robotics will likely determine how trillions of dollars of physical labor are allocated — and most of the popular discourse is getting the shape of that opportunity exactly wrong.
We are at the beginning of a strikingly similar inflection point in robotics. If SaaS decomposed software into modular, composable services, robotics is now decomposing physical labor the same way.
The public conversation, however, is dominated by humanoids.
Atlas does backflips. Optimus folds laundry at half the speed of a tired intern. Figure robots walk through warehouses with the careful deliberateness of someone who has just had a knee replacement. The aesthetic is compelling, the demos are shareable, and the underlying implication — that a robot shaped like a person can step into any role a person occupies — is seductive.
It is also, in most cases, the wrong bet.
“The shape of the robot should be determined by the geometry of the problem — not the geometry of the person doing it today.”
Humanoids Are a Category Error
There is a deep and persistent confusion in robotics between generality and versatility. A humanoid robot is general in form — it can, in principle, operate in any environment built for humans. But the vast majority of commercial problems are not general. They are specific, bounded, and often quite strange.
Consider the task of cleaning the interior of a commercial airliner during its turnaround — the 45-minute window between one flight landing and the next departing. The geometry of the problem is deeply non-human. You need to reach under seats, into overhead bins, across the full width of a row, at precise angles, at speed, with minimal disruption to adjacent tasks.
A humanoid would be slower, clumsier, and less reliable than a purpose-built machine with a compact chassis, multiple specialized end effectors, and sensor arrays calibrated for exactly this environment. The humanoid's arms are too short. Its base is too wide. Its visual processing is tuned for navigating the world, not for finding a forgotten boarding pass under seat 27B.
And critically, a humanoid doesn’t just cost more to build — it costs more to maintain, certify, and deploy. Every degree of freedom is another failure mode. Every additional capability is another system that must be tested across environments it was never optimized for.
The same logic applies across hundreds of verticals. A plumber does not need a robot that can also write emails and play chess. A plumber needs a robot that can navigate a crawlspace, apply precise torque to a fitting in a confined space, inspect a joint with a borescope camera, and detect moisture at millimeter resolution.
The shape of that robot — probably a low-slung chassis with telescoping, multi-jointed arms and a suite of specialized end effectors — bears no resemblance to a human. And it will be vastly better at the task than any humanoid ever will be.
The Long Tail of Specialized Robotics
Here is where the SaaS analogy becomes genuinely illuminating.
In the early years of enterprise software, the conventional wisdom held that the market would consolidate around a small number of massive, general-purpose platforms — SAP, Oracle, Salesforce. And to a significant extent, that happened at the top.
But what nobody fully anticipated was the extraordinary proliferation of point solutions: tools that did exactly one thing, in exactly one vertical, with a depth of domain specificity that no general platform could match.
Robotics is going to develop the same structure.
The addressable universe of physical tasks is vast and highly heterogeneous. Most of these tasks share almost nothing — the geometry, the environment, the required precision, the acceptable failure modes, the economic value of automation are all different. This is not a market that will consolidate around three humanoid platforms. It is a market that will produce hundreds of highly specialized companies, each dominating a specific task environment with a robot purpose-built for that environment.
This is already visible in early deployments: companies building warehouse picking systems optimized for specific SKU distributions, agricultural robots tuned to a single crop type, inspection robots designed exclusively for wind turbines or pipeline interiors.
Vertical example
Aircraft turnaround cleaning
Compact low-profile chassis. Multi-angle articulated arms. Optimized for the 45-minute ground window between flights, reaching under seats and across rows without blocking human crews.
Vertical example
Subsurface plumbing inspection
Telescoping flexible limbs. Borescope cameras. Moisture and pressure sensors. Built for crawlspaces where visibility is near zero and joint configurations vary unpredictably.
Vertical example
Greenhouse pollination
Lightweight, high-precision actuators. Dense computer vision for plant detection. Operating at canopy level during narrow pollination windows without disturbing plant growth.
Vertical example
Bridge cable inspection
Magnetic grip chassis. Long-range structural scanning. Designed for environments that are actively dangerous for humans, operating in wind and variable lighting conditions.
Each of these is a real business. Each has defensible unit economics. None of them require a robot that can make a cup of coffee or fold a shirt.
The Standardization of the Stack
What makes the SaaS analogy stick — and what makes the current moment in robotics genuinely exciting — is what is happening to the underlying components.
In the early days of SaaS, founders had to build nearly everything themselves: their own authentication, their own payment infrastructure, their own email delivery, their own analytics pipelines. The overhead was enormous.
Then, slowly and then all at once, the stack standardized.
Robotics is undergoing a strikingly similar transition.
Five years ago, building a commercial robot meant solving hard problems in motor control, sensor fusion, mechanical design, computer vision, and deployment infrastructure — simultaneously. Today, the commodity layer is forming fast. Actuator packages are configurable and well-documented. Middleware ecosystems have standardized core interfaces. Simulation platforms have dramatically compressed the time to generate synthetic training data.
The implication is that the cost of entering a robotics vertical is falling rapidly.
A small team with genuine domain expertise in, say, the logistics of pharmaceutical cold-chain handling or the specifics of offshore wind turbine inspection can now build a competitive robot without recreating the entire stack from scratch. They adopt the commodity components — actuators, sensor hardware, simulation pipelines, base control software — and focus their energy on the part that matters: deep understanding of the specific task environment.
“A small team with genuine domain expertise can now build a competitive robot without recreating the entire stack from scratch. The commodity layer is forming fast.”
The Real Moat: Computer Vision Data
In SaaS, the canonical moat is data network effects. The more customers a business intelligence tool has, the more benchmarks it can offer. The more transactions flow through a payments platform, the better its fraud models become.
In specialized robotics, the equivalent moat is task-specific computer vision data.
This point deserves emphasis because it is somewhat counterintuitive. The winning robotics companies of the next decade will not necessarily be those with the best mechanical engineers or the most sophisticated control algorithms. They will be those with the deepest, richest, most varied datasets of their specific task environment.
Most robotic failures do not occur in ideal conditions — they occur in edge cases: unusual lighting, partial occlusion, unexpected object configurations, degraded materials, or rare environmental states. The company that has seen 10,000 of these edge cases will outperform the one that has seen 100.
General-purpose vision models will handle the easy cases. They will recognize a cup, parse a sign, understand that a doorknob turns. What they will not do well, at least not for a very long time, is understand the specific failure modes of a specific environment.
They will not know that airliner seat fabric reflects light in a way that creates false positives for a certain class of spill detector. They will not know that the joint configuration a plumbing robot must reach in a 1950s Boston rowhouse is systematically different from the same configuration in a 1990s Houston suburban build.
That knowledge lives in proprietary datasets accumulated through real operational deployment.
The company that deploys robots in airport X and accumulates a year of operational vision data has something a competitor entering that market cannot replicate quickly. This is a genuine, durable, compounding advantage.
Why the Humanoid Is Overrated (With Exceptions)
None of this is to say that humanoid robotics is useless or that the companies pursuing it are foolish.
There are genuine applications — elder care, some classes of domestic assistance, certain military contexts — where the ability to operate in fully unstructured human environments is worth the enormous cost of the humanoid form factor.
But “operate in any environment a human can” is a much smaller set of real commercial opportunities than it appears.
A natural counterargument is that advances in AI will make humanoids sufficiently general to overcome these limitations. That perception may generalize faster than expected.
But embodiment does not.
Even if perception improves dramatically, actuation remains constrained by physics. The physical world imposes limits that software cannot abstract away. Precision, reach, stability, and reliability are all shaped by mechanical design. A system optimized for everything is, in practice, optimized for nothing.
Most commercial environments are not arbitrary. They are highly structured, repeatedly traversed, economically valuable, and amenable to purpose-built solutions.
The factory floor, the warehouse, the aircraft, the sewage main, the solar farm — these are not general environments. They are specific environments with specific geometry, lighting, capabilities, and acceptable failure modes.
The humanoid is the mainframe — powerful, centralized, and expensive. Specialized robots are the microcomputers — cheaper, distributed, and optimized for a single job.
History suggests which one wins in aggregate.
Implications for Founders, Investors, and Incumbents
For founders, the opportunity is to think like a vertical SaaS founder, not like a platform builder.
The question is not “how do we build a robot that can do anything?” It is “what specific task, in what specific environment, has sufficient economic value, sufficient operational repetition, and sufficient geometric consistency to justify a specialized machine?”
For investors, the mental model should shift from evaluating robotics companies on the sophistication of their hardware to evaluating them on the defensibility of their data flywheel.
Which company is accumulating the most irreplaceable operational data in its task environment? Which company’s robots are learning from every deployment in a way that makes the next deployment cheaper and more reliable?
That compounding dynamic — not the elegance of the chassis — is where value accumulates.
For incumbents in labor-intensive industries, the question is not “when will robots replace our workers?” It is “which of our task environments will be automated first, and by whom?”
The Shape of the Revolution
The SaaS revolution did not produce one dominant software company. It produced thousands of specialized ones, each owning a narrow but valuable slice of the market.
The specialized robotics revolution will follow the same path.
It will not look, at first, like the humanoid future that fills the feeds. It will look like a very strange machine doing a very specific job very well in a context most people will never see.
And then, slowly and then all at once, it will be everywhere.