At an Amazon warehouse near Luton, shelving units move as if dancing with automated carriers, but it's still a human worker who pulls products from boxes and stocks the shelves. That detail points to the real bottleneck the AI debate keeps overlooking: while language models have mastered text generation, machines still can't grasp, hold, and manipulate objects with anything close to the skill of a human hand.
What happened?
The Guardian's visit to Amazon's massive distribution center near Luton offers a clear picture of where automation actually stands today. Yellow shelving units swap places with robots, conveyors move products in seconds — but the job of pulling items of different sizes and textures, from a Lego box to a dishwasher tablet to a tub of protein powder, off the shelf and stacking them still falls to humans. Amazon frames this as a "symphony of people and technology," but the reality is this: the robotic arms the company has invested in for years — Sparrow, Cardinal, Proteus — remain limited to rigid, uniform objects.
Why robot hands still aren't as capable as human hands
There's a name for this: Moravec's paradox. Back in the 1980s, roboticist Hans Moravec observed that computers could easily handle tasks that seem "hard," like playing chess, while tasks that seem "simple" — the kind of grasping and balance even a baby can do — turn out to be incredibly difficult to replicate. Forty years later, the paradox still holds. GPT-class models generate language using billions of parameters, but a robotic hand gripping a soft piece of fruit without crushing it, or a slippery glass bowl without dropping it, requires estimating each object's weight, friction, and fragility in real time. The human hand does this with thousands of touch receptors and years of motor learning; for a robot, it means dozens of sensors, actuators, and software layers, each requiring separate calibration.
Boston Dynamics' Atlas robot can do a backflip, but it can't fold socks — because walking and balance are repeatable physics problems, while manual work means encountering a different object every time. Tesla's Optimus and Figure AI's humanoid robots are trying to close this gap, but it has come out that a significant portion of their public demos were actually run via teleoperation (remote control). Figure's pilot program at a BMW factory is also limited to a single part-placement motion — not general-purpose dexterity.
The gap between warehouse automation and humanoid robot promises
This distinction matters because the narrative in investor presentations — "humanoid robots will do every job" — is disconnected from the reality on the warehouse floor.
| Area | Current state | Limitation |
|---|---|---|
| Transport / logistics | Robots move shelves, boxes, pallets | Fixed, predictable objects |
| Selective picking | Partially automated (rigid, uniform products) | Soft, fragile, irregularly shaped products still require humans |
| Humanoid robot demos | Single task in controlled environment | Often teleoperated, not general-purpose |
| Language/decision AI | Advancing rapidly, costs falling | Hard to transfer to the physical world (sim-to-real gap) |
This table shows that AI is changing jobs, but first at the decision-making and planning layer; progress in the physical manipulation layer is far slower. In the near term, the people most at risk of losing their jobs aren't actually manual laborers — they're knowledge workers doing things like route planning, inventory forecasting, and customer service.
What's next?
The real development to watch is the falling cost of tactile sensors and progress in sim-to-real transfer. Research on hands from Google DeepMind and various university labs shows small but genuine steps forward in general-purpose grasping. Also worth watching: how much Amazon, Figure, and Agility Robotics manage to reduce the teleoperation ratio in their warehouse pilots over time — that will be the most concrete indicator of real autonomy.
In the end, the gap between hype and reality remains quite large: humanoid robots won't be replacing the workers stocking warehouse shelves anytime soon. AI's hands lag far behind its language skills, and until that gap closes, "end of humanity" scenarios will remain the stuff of headlines rather than fact.
Yorumlar (0)
Henüz yorum yok. İlk yorumu sen yaz.
Yorum yapmak için Sinyal'i indir
Yorumlar Sinyal hesabıyla yapılır. Mobil uygulamada giriş yap, yorum bırak.