SI Agent News
DevelopingGeminiAgentsResearch

Why scaling AI models won't quickly create autonomous robots

ETH Zurich professor Marco Hutter argues that physical constraints make robotics progress much slower than the rapid scaling seen in Generative AI.

The short version

Why it matters

For those following the AI boom, this clarifies that the 'easy' era of progress via raw compute might be hitting a wall in robotics. It suggests that the next big breakthroughs will require deeper engineering in hardware and physics rather than just bigger datasets.

The limits of scaling laws

The robotics industry is currently racing to apply the same development methods used for Generative AI to physical machines. Companies are attempting to build robot foundation models, such as vision-language-action (VLA) models, to give robots general autonomy. The goal is to create a single model capable of performing diverse tasks across many different types of robots.

However, Marco Hutter, a professor at ETH Zurich and a leading robotics researcher, argues that this approach is overhyped. He notes that while scaling up model size and data worked for digital outputs, it will not improve robots at the same pace. He believes the progress in physical AI will be much slower because robots must operate within the constraints of the real world.

As Chosunbiz reported, Hutter explains that the decisions made by an AI in a physical setting must translate into actual motion. Unlike a digital environment, a robot must move motors and joints to interact with objects. These motions change based on specific physical variables like an object's weight, shape, friction, and contact conditions.

Why physics creates friction

The primary difference between Generative AI and physical AI lies in the necessity of reliability in control and sensing. In a purely digital world, the AI does not have to worry about the mechanics of movement. In the physical world, the AI's intelligence must go hand in hand with the hardware's ability to execute those commands accurately.

Hutter points out that the 'scaling laws' that propelled Large Language Models (LLMs) are difficult to apply directly to robots. He suggests that simply feeding in more behavior data from videos of people working will not solve the underlying challenges of commercialization. The field still requires broad advances in hardware safety and the ability to handle complex contact problems.

Because of these physical constraints, Hutter argues that the industry cannot rely on a single model to solve everything. He believes that the reliability of a robot depends on how well the AI understands the specific physics of its environment. This means that the path to general-purpose robots is much more complex than just increasing computing power.

The path to reliable skills

To overcome these hurdles, Hutter suggests that the biggest progress will come from combining robot foundation models with specific learning techniques. He highlights Reinforcement Learning (RL) as a key method, which allows robots to learn behaviors by repeating trial and error to maximize rewards. He also emphasizes the importance of large-scale simulation as a training ground for these skills.

Hutter believes that learning physical skills through simulation and RL remains underrated in the current landscape. He also points to force and tactile sensing as essential technologies for reliable interaction. These tools help robots address the difficult contact and control problems that occur when a machine tries to manipulate an object in 3D space.

Furthermore, he identifies 'agentic behavior' as a major direction for the next year. This involves robots planning complex, long-horizon tasks, tracking their own progress, and responding when problems arise. When combined with tactile sensing and fine manipulation skills, these capabilities can turn abstract plans into reliable physical work.

Hardware vs. Intelligence

There is a growing assumption in the tech world that as AI becomes more advanced, the differences between various robot bodies will narrow. The idea is that the same high-level intelligence will eventually be able to run on almost any type of robot hardware. Hutter challenges this assumption, stating that a robot's physical design determines what it can sense, learn, and do.

He argues that machines should be designed specifically around the tasks they are intended to perform. Instead of trying to create a universal body for a universal brain, he suggests that the most powerful products will come from developing hardware and intelligence together. This means the physical form of the robot must be optimized for its specific job.

By focusing on the synergy between physical design and AI, researchers can create more effective tools for specific industries. While the push for general-purpose robots continues, Hutter's perspective suggests that specialized, hardware-integrated intelligence will be the more practical path forward. This approach prioritizes the reality of physics over the convenience of digital scaling.

The road ahead for robotics

The robotics industry is currently at a crossroads between the

Sources and links

Liked it? Say why

Add a line for other readers.

No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.

Listen to the AI Wire headlines

Listen to the AI Wire

Listen to our daily AI shows every morning: on your commute, in the shower, wherever you are

Listen to the daily shows

First reported by Chosunbiz · Read the original. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.

More AI news