When will Bicentennial Man or Ava move in? The race to build synthetic humans

A familiar scene is beginning to appear across street markets, factory floors, and workshops around the world. Workers sorting produce, stitching cloth, or assembling components are increasingly seen wearing sensor rigs, head-mounted cameras, and motion-capture bands while going about their daily tasks. To most people, these devices may look unusual. To the AI and robotics industry, they are valuable.

The movement data gathered from these workers is being used to teach machines how to grasp, handle, and manipulate objects in the real world. These are not merely gadgets strapped onto human bodies. They are real-world data collection systems, increasingly important for the next phase of robotics development.

That data is being collected because the technology industry is steadily pursuing an ambitious goal: the synthetic human, reminiscent of Andrew Martin from Bicentennial Man or Ava from Ex Machina. What once belonged to science fiction is now firmly on engineering drawing boards, with early versions already entering the market. Machines that can move, perceive, decide, and respond with something close to human presence are no longer confined to films and novels. Building them has become a serious ambition for some of the world’s largest technology companies.

More from Opinion

The investment is enormous, and so is the excitement. But the timeline is likely to be far longer than many technology leaders and AI optimists expect.

Andrew Martins may eventually emerge, but only after many iterations and many years, if not decades. Two difficult problems remain unresolved: reliability in AI software, the machine’s brain, and seamless dexterity in robotics hardware, the machine’s body.

The Problem of Trust

The first challenge is reliability.

Today’s AI can draft, summarise, diagnose, and reason in ways that often appear remarkably human. But it still hallucinates. It can produce confident but incorrect answers without warning or explanation. OpenAI itself has acknowledged hallucinations as a persistent problem in language models.

That becomes a serious obstacle when AI moves from chat windows into physical environments. A synthetic human cannot function correctly only “most of the time". It must operate with predictable reliability, especially in situations where mistakes could injure someone.

That is the real test.

A machine that performs brilliantly during a controlled demonstration but behaves unpredictably in a crowded corridor, a hospital ward, or a busy kitchen is not ready for the real world. In human environments, small errors can create serious consequences. Reliability is not an optional upgrade. It is the foundation.

The Problem of the Body

The second challenge is dexterity.

Human work is not simply movement. It is movement guided by touch, pressure, balance, timing, and constant adjustment. A vegetable vendor sorting produce, a tailor stitching fabric, or a surgeon closing an incision is not merely following a sequence of actions. The body continuously reads the environment and adapts in real time.

The human hand alone has twenty-seven degrees of freedom (DOF). It operates through an intricate feedback system refined over years of practice and lived experience. Reproducing that capability in a machine is not a minor engineering problem. It is an entirely different level of difficulty.

Robots can already move with impressive speed and accuracy. But human dexterity is more than motion. It requires sensing, adapting, correcting, and recovering under uncertainty.

That is why the workers wearing sensor bands matter. They are not only producing goods. They are also generating movement traces, visual frames, and behavioural data that may help train the next generation of robots.

When a machine watches a person fold cloth or picks up a fragile object, it is learning relationships between vision, touch, timing, and correction. The Open X-Embodiment collaboration reflects the scale of this effort, pooling more than one million robot trajectories from twenty-two robots across twenty-one research institutions to build RT-X models capable of transferring skills across platforms.

That is meaningful progress. It is also a reminder of how much work remains before machines can approach what humans learn naturally through everyday life.

Why Humans Are Hard to Copy

The deeper problem is evolutionary.

Evolution did not build the human mind first and the body later. Cognition and physical skill developed together over hundreds of millions of years, constantly shaping each other through interaction with the environment. Human intelligence emerged through movement, sensation, adaptation, and survival in the physical world.

A synthetic human built from an AI brain and a robotic body does not inherit that long evolutionary process. Instead, it combines two highly advanced systems developed separately and then attempts to make them function as one.

Science fiction has often understood this problem better than technology marketing.

Ava in Ex Machina appears fluid and convincing, yet still feels alien underneath. The Hosts in Westworld are physically impressive, but their creators repeatedly fail to fully align cognition, memory, emotion, and embodiment. These stories return to the same conclusion: human presence cannot simply be assembled like machinery. It develops through long interaction with the world.

Hans Moravec identified this paradox decades ago. Tasks humans find intellectually difficult, such as chess or complex calculations, can become relatively easy for computers. But tasks humans perform effortlessly, such as catching a falling object, reading social cues, or handling fragile materials, remain extraordinarily difficult for machines.

AI has overturned this pattern in some narrow areas. It has defeated world champions at Go and predicted protein structures that challenged scientists for decades. But the sensorimotor layer of human intelligence, shaped through deep interaction with the physical world, remains stubbornly difficult to replicate.

Where the Field Is Heading

Many of the field’s most respected researchers are now focused on this exact gap.

Fei-Fei Li is working on what she calls spatial intelligence: the ability for machines to perceive, reason about, and interact with the three-dimensional physical world, not merely process language. Her work aims to give AI systems a stronger grasp of physical reality, including weight, distance, movement, and consequence.

Yann LeCun has argued that current AI architectures are insufficient for achieving human-level intelligence. His proposed direction focuses on helping machines learn underlying models of how the world works, rather than simply identifying patterns in text.

Together, these efforts point toward the same conclusion: the future of advanced AI will require systems grounded in the physical world, not just language.

A machine that speaks fluently is not necessarily a machine that understands reality.

A Bridge, Not the Destination

The synthetic human is not impossible. Given the current pace of technological development, some version of it may eventually emerge. But public discussion often confuses the easier challenge, making robots look human, with the far harder one: building systems that can reliably combine perception, judgement, movement, and trustworthiness in unpredictable real-world environments.

The workers wearing sensor rigs today may help train the robots of tomorrow. But this is still a bridge, not the destination.

We may eventually see Andrew Martin and Ava among us. Just not soon. Patience is not pessimism.

(Sreejith Sreedharan is a technology analyst and author of Future of Work – AI-Augmented Autonomous Decentralised. He works on organisational AI readiness and created the AI Instinct Index®, a psychometric diagnostic designed to assess behavioural readiness for AI adoption and adaptive capacity in constraint-heavy environments. Views expressed in the above piece are personal and solely those of the author. They do not necessarily reflect Firstpost’s views.)