We are currently seeing a surge in the field of robotics, with dozens of companies developing highly complex systems – some of which already demonstrate remarkable feats such as parkour running, backflips, or performing kung fu–like movements. These performances often surpass the agility of many of their human creators. They are, without doubt, impressive technological achievements, built on decades of research in mechanical engineering, sensing technologies, material science, control algorithms, physics simulation – and, of course, machine learning.
There is only one problem.
Kung fu moves make for a great show and, quite frankly, attract significant investment from those eager not to miss the next big opportunity in a rapidly growing market. But companies do not purchase robotic systems based on impressiveness alone. Whether a robot becomes a valuable addition to an organization ultimately comes down to one simple question: What is its return on investment?
Answering this question introduces very different considerations. How much do existing processes need to change to integrate the robot? Can it perform its task effectively and reliably? How much effort and cost are required to keep it running? How often does it fail – and when it does, how long is the downtime? And perhaps most importantly: what can this robot actually do?
These problems seem obvious – but how do we begin to address them?
A useful place to start is with the basics. Historically, major technological leaps have emerged within the realities of their time, building on existing infrastructures and methods. The automobile spread because road networks already existed for horse-drawn wagons; early electrical grids followed gas-lighting networks; and the internet initially ran over telephone lines. In each case, new technologies gained traction by fitting into familiar structures before eventually reshaping them.
The same premise must hold for robotics if it is to succeed. We do not want to reshape the world to accommodate the robots of the future; instead, robots must be designed to fit into our world – right here, right now. And right now, the world is shaped for humans. Every door handle, every cup, every tool – almost everything around us – is made by us, for us, to be used by us. These objects allow humans to interact with the world effectively: to experience it, operate within it, and shape it.
This principle extends far beyond everyday objects. It is even more pronounced in manual labor and industrial environments, where tools, workflows, safety standards, and workspaces have been optimized around human bodies, perception, and decision-making. Any robotic system meant to operate in these contexts must therefore adapt to a human-centered world, rather than expect that world to adapt to it. Simply put: to fit into a human world, robots must be human-like – at least in the ways that matter.
But here come the counterarguments, grounded in what appears to be the strongest possible evidence: reality itself. Hasn’t the history of industrialization shown that success is driven by fine-grained specialization, deliberate division of labor, efficient resource distribution, highly specialized automated machinery, and mass production? From food processing to car assembly, productivity gains have traditionally come from narrowing scope, simplifying environments, and designing machines to handle clearly defined tasks with maximum efficiency.
And indeed, this approach works. Decades of automation have dramatically reduced the need for human labor in manufacturing. In the process, we have built the robots of today – and they look nothing like humans. So what, then, is the point?
The point is what this argument leaves out: everything else.
In developed economies, most large-scale industrial infrastructure is already in place. The growth model of traditional industrialization has largely reached its limits. Factories are built, processes optimized, machines deployed. Much of what proved too difficult to automate was not solved, but outsourced to regions where manual labor is far cheaper than domestic labor. Case closed. Nothing left to do. Perfect. Time for lifelong vacations, cocktails, and beaches.
Or rather not.
In reality, there is plenty of work left – just not inside the well-organized, highly structured world of factories and assembly lines. What remains are jobs in the real world: maintenance, construction, logistics, care, and other forms of manual labor embedded in human-centered, highly variable environments. These are precisely the domains where further productivity gains cannot be achieved through ever more optimized assembly lines, but instead require systems capable of operating in the world as it exists outside the factory – diverse, dynamic, and often chaotic. In other words, we need robots that can handle chaos. Achieving this demands extreme levels of versatility – and one of the most promising paths toward such versatility is to follow the human example.
But do robots really need to look exactly like humans?
Probably not. And most likely, most of them won’t – at least for the majority of tasks. In many cases, a full human-like body with two arms, two legs, a torso, and a head would simply be overkill. A robot on wheels, for example, is perfectly adequate for most logistics environments. Claiming that such a robot is useless because it cannot climb stairs is like arguing that a forklift is useless for the same reason. Both are effective precisely because we deploy them in environments already adapted to their capabilities.
So which aspects of a robot, if any, need to be human-like?
As you might expect, the answer depends on the task. There is a critical balance to strike between simplicity and versatility. Designing a robot for a narrowly defined task enables specialized hardware and often results in systems far simpler than the human form. At the other extreme, humanoid designs offer broad versatility by directly exploiting the human-centered structure of our world – tools, spaces, and workflows shaped around the human body.
The central challenge in robot design, then, is not to copy humans wholesale, but to decide which human capabilities are essential – and which are unnecessary – for a given application.
There is, however, one aspect of the human body so central, so versatile, and so universally applicable that it makes sense to deploy it in most robotic applications operating in human-centered environments. Humans rely on it in almost every situation, every day, everywhere. Through it, we access the environment, interact with it, shape it, and build it. Countless tools, workflows, and environments have been designed around its capabilities and shaped to its form. The world, quite literally, is made to work with it.
It is nothing less than the interface between the inner mind and the outside world.
That interface is the human hand.
Seen through this lens, the human body can be understood as a system that carries the mind and empowers the hands that act on its behalf. But let’s not get too philosophical.
Do robots really need full-blown, human-like hands with five fingers?
Most likely not. And to make matters worse, developing a robotic hand with human-level capabilities is widely regarded as one of the most challenging problems in robotics. It is entirely conceivable that future hand designs will be discovered that are far simpler than the human form, yet offer comparable – or even superior – levels of dexterity, versatility, and reliability. After all, the human hand represents only the current stage of an ongoing evolutionary optimization process – one that may not have converged yet. It would be naive to assume it represents an ultimate optimum beyond which no better design can exist.
We therefore face an uncomfortable reality: we are attempting to replicate a mechanical system whose optimality is unknown, while simultaneously pursuing one of the hardest technological challenges in the field.
But here is the catch: by making the mechanical problem harder, we can significantly simplify the application problem.
And this is a trade-off worth embracing. Building a robot is challenging – but operating it reliably in the real world is often just as difficult. With recent advances in artificial intelligence, this challenge appears more tractable than ever. The caveat, however, is that modern AI approaches require large amounts of training data – data that is difficult to obtain and has traditionally relied on simulation.
By using hardware that closely mimics the structure and function of the human hand, we gain access to a powerful new source of training data: human demonstration. This allows us to directly transfer human strategies to robotic systems without the need to infer or translate observed behaviors – or even worse, to engineer control strategies for each task separately.
Learning from human demonstration enables a radical shift in how we interact with robots. The robots of the future will not be programmed like machines. They will be taught like apprentices.
Conclusion: Dexterity as an Economic Lever
Ultimately, the question of what makes a good robot is not one of aesthetics or impressiveness, but of economics. A capable robot is defined by how effectively it interacts with its environment and by how easily it can be integrated, trained, operated, and maintained within real-world processes.
And as the primary means of interaction, the hand becomes more than a mechanical component – it becomes an economic lever. While human-like hands are mechanically complex and difficult to engineer, they dramatically reduce complexity elsewhere: in application design, task specification, environment adaptation, and data collection. By aligning robotic hardware with the way humans already interact with the world, we lower the cost of deployment, reduce the need for specialized infrastructure, and unlock the potential of fast, intuitive, and powerful learning through human demonstration.
Seen through this lens, human-like robotic hands are nothing but a pragmatic response to the economic realities of automation in a human world.
