Today’s humanoids perform circus tricks and karate kicks. Their human-like movements are fascinating, but they are only rehearsed routines in a prepared environment, and they require vast amounts of recorded human data to work. Change something in the environment and they risk getting it wrong.More data doesn’t solve the problem
The most common misconception is that more data is the key to making humanoids succeed, but more hours of human teleoperation aren’t enough. Physical work in production is highly varied, and you’re constantly faced with new situations. A problem can be solved in several ways. You can move a box either by lifting it or by pushing it. That’s obvious to humans but not to humanoids. Solving real-world problems requires an understanding of the world and the ability to reason in multiple steps. That is missing from today’s humanoid brains, which are driven by hallucinating language models.
A world model instead of a language modelWhat’s needed is an AI brain with a world model: a map or representation of how the physical world behaves, one that can predict the consequences of an action the robot has never performed before. The robot must understand the properties of objects, such as weight, friction and balance, and be able to handle order, for example what lies beneath the top item in a crate. Without a structured world model, every object and every task requires new AI training. That is neither scalable nor sustainable. The cost of AI training has skyrocketed, and the largest language models cost over a billion dollars to train.
The human body is a special case of a mobile manipulatorThe human body is an answer to environmental adaptation: working in an environment built for humans, with stairs, handles and tools with grips. That’s a reasonable argument for the humanoid form factor, but a mobile manipulator like R2-D2 in Star Wars is often a far better design for industry. It has wheels instead of legs, a low center of gravity, no balance issues, and a hatch full of tools that fold out when the task demands it, not just hands.
It’s the ability to use tools that separates humans from chimpanzees, not the handA hand is an excellent gripper, but in many industrial cases it’s the wrong tool. A five-fingered hand is also expensive and fragile. Suction cups beat hands for picking a wide range of objects such as cartons or bags. Magnets work better for metal sheets. Humans choose tools based on the task and know which tool works best. A robot must be able to do the same. That’s why a tool changer is more important than a hand.
Robotics’ simplest task is unsolvedPick-and-place, moving an object from A to B, is an unsolved problem in robotics and far harder than karate kicks, which are really just a balance problem. That may sound like an exaggeration until you grasp the complexity of the seemingly simple. Moving arbitrary objects (soft, shiny, wrinkled, semi-transparent, tangled) in every kind of environment, whether a factory, a forest or a messy home, and placing them in a dresser drawer, a closed screw-top jar or on a clothes hanger requires real intelligence and an understanding of the world: what an object is and what happens when you touch it. Humanoids may well arrive, but the path doesn’t run through more data or more human-like mechanics. The missing link is the brain.
“SICS.AI has developed a structured world model that reduces AI hallucinations by more than 99 percent and cuts computing costs by up to 90 percent. We control industrial robots to perform close to 100 percent correctly in complex environments,” says Karim Nouira, CEO of SICS.AI.
