AI is what Makes Humanoid Robots possible

Humanoid robots would be far less interesting without the current development in artificial intelligence. Motors, cameras, sensors, batteries, and mechanical structures have existed for a long time. AI changes the ability to combine these components into a machine that doesn’t simply repeat programmed movements but can increasingly understand its environment and decide how to act within it.

That distinction separates the humanoid robot from conventional industrial automation. A traditional robot is normally programmed for a clearly defined process and performs that process extremely efficiently as long as conditions remain unchanged. A humanoid is intended to become a generalist. Instead of programming every movement individually, the long-term goal is to give the robot a task and let the system determine how to perform it.

Physical AI turns Information into Movement

The term Physical AI describes this combination of artificial intelligence and hardware. Whereas AI systems on a computer can perceive information, reason about it, and produce an answer, robotics adds another step. Vision-Language-Action models do not produce words on a screen. Their output becomes a physical action executed through motors and actuators.

That difference is fundamental. If an AI model produces an incorrect sentence on a computer screen, the consequences are limited. If it controls a robot carrying ten kilograms through a factory, its decisions immediately interact with people, equipment, and the surrounding environment. The system’s intelligence must therefore be combined with functional safety precautions before it becomes industrially useful.

Training is still more Complicated than Prompting ChatGPT

The long-term vision is easy to understand. You show a robot what needs to be done, provide some background information, and tell it to perform the task. The current training process is still far away from that simplicity.

Humanoid currently starts by identifying the process that needs to be automated. The company then recreates a digital twin of the customer’s environment with comparable components, tables, or conveyor belts. Data collectors operate the robot through teleoperation and repeatedly demonstrate the required action. The resulting data must be processed, labeled, and used for pre-training and post-training before the robot can eventually perform the task autonomously.

This makes training both labor- and compute-intensive. However, the advantage appears once this knowledge becomes transferable. A conventional automation project requires engineering effort for every new task. An AI model can be deployed to many robots and eventually even across different hardware generations. The work required to teach the first robot therefore has the potential to create value far beyond that individual machine.

Methods such as reinforcement learning will push this further. Instead of relying so heavily on human demonstrations for every individual action, future robots should increasingly be able to improve through their own interaction with the environment. This is an important step toward making training easier and allowing one software model to work across several hardware versions.

The Future will look similar to today’s LLMs

The logical end point is a training process that feels much closer to how we already interact with generative AI. Instead of programming trajectories, the user could provide instructions, images, videos, and background information and tell the robot what outcome is required.

The robot would then use that information to perform the task. A production employee would no longer need to become a robotics programmer simply because another product has to be moved from one box to another. The robots solve the complexity with their own AI model inside.

That could dramatically change the economics of automation. Today, one of the biggest barriers to automation is not necessarily the price of the robot itself but everything required to make it useful for one specific application. If training becomes progressively easier, the same hardware will suddenly be used for far more tasks.

One Factory becomes One Learning System

A humanoid robot rarely comes alone. Normally, several humanoids operate as a fleet, with time-critical control and short-term decisions processed inside each robot while higher-level coordination happens externally.

A central system assigns tasks, coordinates routes, and decides when individual robots should recharge. In a manufacturing environment, this becomes increasingly complex because robots must also coexist with employees, machines, and internal logistics. In a foundry, that environment also includes forklifts transporting liquid metal or castings being moved between operations.

The interesting consequence is that knowledge gained by one robot does not necessarily have to be limited to that robot. Once a task has been trained and validated, the software can then distribute it throughout the fleet. And the value of the robots rises with the data and intelligence available to the entire system.

The Training Data is Intellectual Property

A humanoid robot is equipped with cameras and other sensors that observe its surroundings. Training it inside a factory therefore will generate images and videos of machines, layouts, castings, and processes that would normally be protected through NDAs.

Who owns that information, where it is processed, and who is allowed to use it consequently becomes an important purchasing question. According to Wilberth, this area is still developing. For now, contractual protection of the collected data is required. Working with suppliers operating under the same European or North American data regulations can simplify this problem.

In the longer term, the solution is to move more of the training capability to the customer. If a foundry can teach its robots locally without continuously sending production data to the robot manufacturer, both the usability and the data-security problem become easier to manage.

AI is therefore not simply another feature added to the humanoid robot. It is the technology that makes the entire concept scalable. The hardware gives the robot its physical capability, but the software determines how many different jobs that hardware can ultimately perform.