A Tübingen World Model as the Spearhead of AI
The startup Ontic Labs is working on a training program designed to bring robots up to ChatGPT level.
While language models like ChatGPT or Claude are now capable of providing a mathematical proof that has stumped humanity for over 30 years, the leaps in robot development have been rather modest. They still primarily perform monotonous tasks in factories, rather than taking on more complex work in our homes, on construction sites, in hospitals, and in public spaces. The reason is simple: the availability of digital data makes all the difference. Language models have access to a nearly inexhaustible source of texts in every type and form. In contrast, the supply of physical data—the kind needed to train machines—is in short supply. Five founders from Swabia’s Cyber Valley have set themselves an ambitious goal: With their startup Ontic Labs, they aim to pave the way for a “ChatGPT for robots.” The Tübingen-based scientists are working on a training model that—in a cost-effective and universally applicable way—is intended to catapult robotics to the level where language models are today. Wieland Brendel and Omid Taheri from the Max Planck Institute for Intelligent Systems, as well as Georg Martius, Gerard Pons-Moll, and Mikel Zhobro from the University of Tübingen, are relying on videos as a means of conveying physical knowledge.
Traditional Robotics and AI World Models
Traditional robotics relies heavily on rule-based programming or patterns learned from limited datasets. This works exceptionally well in stable and simple environments, but quickly fails when faced with unpredictability: lighting conditions change, objects are positioned differently, and people move dynamically. That’s why, for example, unloading a dishwasher turns into a fiasco when a robot is tasked with this task—which is very simple for a human but highly complex for the robot. The use of so-called world models promises a solution. They change the basic principle: Instead of merely reacting to commands, an AI system learns to simulate the physical and causal relationships in its environment. It develops an internal “understanding” of what happens when an object is moved, bumped, or touched by a person. Another advantage: scenarios can be simulated without damaging hardware or endangering people. However, the training effort required with the simulation technology used so far is immense—and therefore expensive. Ontic Labs has set out to change that.
Understanding instead of Imitation
Videos provide a wealth of physical data and are already being used extensively in robotics. However, their application has so far been limited to the simple simulations of motion sequences described above, which robots imitate without understanding them. Robots trained using Ontic Labs’ world model ultimately come to understand physical relationships. Machines can then also handle situations they have not yet encountered. This would bring Ontic Labs very close to a “ChatGPT for robotics”: any problem could be solved without additional, costly training. With a general understanding of physics, robots gain the ability to move independently in this world and take on complex tasks. And as generalists, they would then be versatile—including the ability to unload that dishwasher without causing any accidents.
“If a robot understands how its actions change the world, it can act safely, efficiently, and adaptively,” confirms Brendel. “Videos provide our world model not only with visual patterns but also with physical laws, friction, inertia, and human behavior.” While today’s AI models primarily capture what the world looks like, Ontic Labs’ technology will understand how it works—thereby enabling robots to interact more complexly with their environment. “We’re developing a basic model that learns from videos and experience how the world works physically. Our goal is to train robots using this basic model,” says Georg Martius, co-founder of Ontic Labs. “This will lay the foundation for more versatile robotic systems that can be deployed in a wide variety of fields in the future.” If their world model succeeds, any robot application could be trained quickly and cost-effectively—a milestone in AI development. With this key technology, the startup would make an important contribution to Germany’s high-tech agenda and to Europe’s digital sovereignty, as Europe seeks to become less dependent on the world’s dominant technology providers.
Startup Sprint for Ontic Labs
With its approach to physical AI, Ontic Labs won the first round of the Next Frontier AI Challenge organized by the Federal Agency for Breakthrough Innovation (SPRIND) this summer. They are among a total of ten teams selected from 550 applicants across Europe to develop fundamental AI technologies with exceptionally high innovation potential. The goal of the challenge: to build three European AI startups within 24 months that, after two additional stages of the competition and successful funding, can achieve a valuation in the billions. If all goes according to plan, Ontic Labs will be among them.
The transformation from an agile research group into a business-oriented startup quintet happened quickly: After submitting their application in May, they received confirmation in June that they had been accepted—on the condition that they establish a startup within a week. Ontic Labs GmbH became a reality. “That was a tight squeeze,” recalls co-founder Brendel. “Just securing an appointment with a notary proved to be a challenge. Registering with the commercial registry, opening a bank account, obtaining a tax ID—the organizational aspects of starting a company can sometimes be a grueling undertaking,” says Brendel, who has extensive startup experience. He and Martius had previously worked on Polybot, an intelligent agricultural robot developed at the Max Planck Institute for Intelligent Systems. “Entrepreneurship needs this kind of speed at SPRIND. Especially with today’s development cycles, every delay is a competitive disadvantage. And founders always want to get started right away.” “Just do it” is also his advice to those interested in starting a business.
