Why Physical AI Is Harder Than We Thought

Physical AI is rapidly moving from impressive demonstrations to engineering reality. If attention was once focused primarily on its capabilities, today the question of scalability is becoming increasingly urgent: what is preventing these systems from becoming truly widespread and reliable?

Physical AI and humanoid robotics now stand at the intersection of three major challenges – engineering, cognitive, and investment-related. Intelligence operating in the physical world imposes fundamentally different requirements than software-based AI: here, mistakes are costly, and the environment remains unpredictable. That is why the conversation is shifting from the wow effect to concrete technical, market, and regulatory barriers.

Mechanics who must learn to think

The first challenge is fine motor skills. We have motors and servos capable of executing highly precise micro-movements. But replicating human sensitivity, flexibility, and the ability to adapt instantly when handling small objects is extraordinarily difficult. The human hand unconsciously regulates force, angle, speed, and trajectory – all within fractions of a second, constantly adjusting to even the slightest changes.

The second challenge is balance and force control. A robot must interact with objects of different shapes, weights, and textures: an apple, a glass, a piece of jewelry, a metal component, a wet or slippery item. A robot may possess significant physical strength, but it must be able to calculate and apply that strength correctly. This requires tactile sensors, systems that allow it to “feel” pressure, resistance, and surface contact. Equally important is not just detecting force, but interpreting it properly within the context of a specific action. It becomes a question of understanding the physical properties of objects – material resistance, elasticity, friction, and other parameters.

Another serious challenge is spatial orientation – the so-called 6D representation. This does not refer to a “six-dimensional world” in a science fiction sense, rather to three positional coordinates, height, width, and depth, plus three orientation coordinates: the angles of rotation along each axis. For example, a tube or a glass is a three-dimensional object. But for a robot, knowing its coordinates is not enough. It must understand the object’s orientation, its position relative to gravity, and how its position will change as the manipulator rotates. If a robot picks up a glass and wants to pour water from it, it cannot simply “tilt the object.” It must calculate the precise trajectory, angle, and rotation speed, while accounting for the liquid inside, its inertia, and the force of gravity. All of this requires sophisticated spatial modeling and prediction of the consequences of action.

Why the market is still cautious

When considering physical AI in the context of humanoid robotics, it is important to acknowledge the still noticeable level of skepticism.

Part of this skepticism is psychological. The uncanny valley effect – when something appears almost human but not quite realistic enough – creates discomfort and anxiety. Unnatural facial expressions, slightly rigid or “broken” movements, mechanical intonation – all of this generates emotional resistance. And technologies that evoke unease tend to be adopted more slowly.

But the main barrier is economic. Investors see that companies have been showcasing impressive prototypes for decades, yet scalable commercial models remain limited. Technological progress is evident, but a sustainable mass market has not yet fully emerged.

Players like Boston Dynamics build engineering masterpieces, yet their applications remain niche and expensive. Tesla is developing its own humanoid projects. New companies such as Figure AI are attracting significant investment, promising robots for manufacturing, logistics, and care industries.

Manufacturing remains an obvious direction in this context. Robotization, there is no question of if, but of speed and cost of deployment.

An even clearer example is logistics and warehousing. Logistics robots are already among the most profitable and widely adopted segments of robotics today. I remember that, at Keymakr, many logistics companies approached us for annotation services while implementing such technologies, with ambitious plans for scaling them further. The scale of global e-commerce demands the movement of enormous volumes of goods at high speed and precision. Humans are physically incapable of operating at that pace. As a result, warehouse automation has become a “hot” topic, giving rise to an entire industry: autonomous platforms navigate routes, sort, transport, and distribute cargo.

Nevertheless, much of the industry remains in the pilot phase and is making ambitious promises. Companies are still searching for compelling use cases that deliver predictable monetization. Investors, in turn, evaluate time to return on investment, technological risks, and the scale of engineering challenges.

That is why the market is developing incrementally. Capital in this field requires not only vision, but proven economics.

Risk becomes part of the architecture

A separate layer of discussion concerns regulation and cybersecurity. A comprehensive regulatory framework for physical AI has not yet been fully formed. The industry is still in its formative stage: there are no mature standards, no widespread presence in everyday environments, and no established certification protocols. Regulations will inevitably emerge – but, as in other technological cycles, they will be a consequence of scaling.

What matters even more right now is a different question – trust in systems that gain physical autonomy. A robot in a home, warehouse, or critical infrastructure facility is a network node equipped with sensors, cameras, microphones, and communication channels. Its behavior is determined by software and updates. And even if a robot is initially programmed to perform only safe actions, the possibility of cyber threats remains. With insufficient protection, malicious actors could theoretically gain access to a network of devices and attempt to use them for harmful purposes.

Scenarios involving the hacking of autonomous vehicles or robot networks are already in the cards. They are treated as part of risk assessment – much like what once happened with banking systems, the internet, and cloud services.

But history shows that technological progress rarely stops because of threats. Instead, industries strengthen protection by establishing standards, implementing monitoring, and building multi-layered security systems. Physical AI will follow the same path. The question is not whether risks will arise, but how quickly security becomes embedded in the entire ecosystem.

An industry is being built around it

All the challenges mentioned, technical, market, and regulatory, share one important characteristic: none of them can be solved in isolation.

Physical AI cannot be viewed as a standalone product or even as a single technology. What we are witnessing is the formation of an entire infrastructure in which hardware, computing, energy, data, and materials evolve in sync. And it is precisely here that it becomes clear: this is the emergence of a new industrial ecosystem.

A robot is autonomous and mobile. This means it cannot rely solely on the cloud. Unlike LLMs running on server clusters, physical intelligence must make decisions locally, in real time. This fundamentally changes chip requirements: they must be powerful, energy-efficient, and optimized for inference on edge devices.

This, in turn, creates a broad spectrum of new development areas: energy-efficient chips for robotics; compact, optimized AI models for edge deployment; platforms for training such models; data annotation systems and the preparation of specialized datasets like what we do at Introspector, as well as advances in batteries and autonomous power systems.

Concepts are already being discussed for a robot to replace its own batteries: removing a depleted module, placing it on charge, and connecting a charged one without fully shutting down the system. This alone could become a separate market.

A comprehensive industry is gradually taking shape around physical AI. Aside from computing and energy, materials science will need to evolve: synthetic coatings that mimic skin, flexible sensor surfaces, safe and tactilely pleasant materials for human interaction. If a robot operates alongside people, its appearance and physical characteristics become part of user perception and trust in the technology.

In this sense, physical AI is about the entire technological stack, from chips and batteries to sensors, software, materials, and human perception factors. It is within this complexity that the true scale of the future industry lies.