Representatives from Mercedes, Siemens, Beckhoff, Arburg, Hawe, KSB and Sick, among others, spent three days in Zug am Arlberg discussing their Industrial AI strategies at the invitation of HANNOVER MESSE and the Industrial AI Podcast. New collaborations and ideas were developed in three working groups.
One major topic was the interplay between robotics and AI and, in particular, the questions of what are world robotics models and how can such models help us? There are gaps in definitions, but Prof. Dr. Stratis Gavves from the University of Amsterdam was on hand to answer questions from the participants in the working groups. He and his team at the chair are working intensively on such models. His answer: World Robotics Models enable robots to learn and use an internal representation of the world in order to act intelligently, safely and adaptively. They are central to the further development of robotics in industry. In keeping with the theme, the drone start-up Spleenlab from Jena gave a live demo of its reasoning drone based on xLSTM technology. This also attracted the local mountain rescue service and the Vorarlberg police.
The second major topic was the revival of reinforcement learning. The topic was represented by Jan Koutnik, who researched the subject for many years together with Prof. Dr. Jürgen Schmidhuber and has since implemented numerous reinforcement applications in industry with his company Evoptima. Perhaps the most important reason for the revival is the integration of reinforcement learning with large pre-trained models such as ChatGPT or Gemini. RL can now also be trained efficiently in robotics. Today, RL is increasingly being used for tasks where classic control logic fails - because the systems are too dynamic, too uncertain or too complex.
The third topic at AI in the Alps is traditionally the Industrial AI Use Cases, which are thematically supervised by Prof. Dr. Marco Huber from Fraunhofer IPA in Stuttgart. A major topic there was agentic applications, but also “good old fashioned AI” approaches. This year, the working group was expanded to include the topic of UX design and industrial AI use cases, because there is a great need in the industry for explainability, comprehensibility and therefore good UX design. Tom Cadera from CaderaDesign was the contact person and presented a new generation of user interfaces.
According to Gartner analysts, more than 40% of agent-based AI projects will be canceled by the end of 2027 due to escalating costs, unclear business value or insufficient risk controls. Most agent-based AI projects are currently early-stage experiments or proof-of-concepts, usually driven by hype and often misapplied, they say.
A new study from UIUC, Stanford University, the University of California, Berkeley, and other institutions examines the benchmarks we use to evaluate AI agents – and the results are troubling. What if these benchmarks are flawed even more than we thought? We found a very interesting paper on this topic and read it immediately.
World robotic models are becoming increasingly important, and one technology plays a special role in this: Gaussian splatting. It is a 3D rendering technique in which scenes are represented by a set of 3D Gaussian distributions, or "splats," rather than by traditional mesh-based or volumetric representations. Our video tip demonstrates how it works.
Many researchers no longer speak of AI models, but rather, AI systems. Will LLMs become the new operating systems in the future? We spoke to Richard Socher about this topic. The German wunderkind is the CEO of You.com in Silicon Valley and recently visited Germany for a few days. We took advantage of the opportunity. Our listening tip.
