From specialized AI systems to adaptive AI systems

How can AI systems successfully apply learned behaviors to new or previously unknown situations? This is the question that Dr. Andre Biedencap will address in future research at the Karlsruhe Institute of Technology (KIT). Biedenkapp has secured funding from the German Research Foundation (DFG) to the Emmy Noether Group for research on improving the generalizability of reinforcement learning methods and will receive €1.2 million over the next three years. DFG has committed to provide additional funding of EUR 920,000 over a further three years, subject to a successful interim evaluation after the initial period.

“The Emmy Noether Award helps outstanding young researchers gain academic independence early in their careers,” said Professor Stefan Hintz, Vice-Chancellor Early Career Research Fellow at KIT. “Dr. Andre Biedencap is a prime example of an outstanding young researcher at KIT for his work on developing adaptive AI systems.”

Reinforcement learning (RL) is an artificial intelligence (AI) learning paradigm in which an AI agent learns how to behave in a particular environment through trial and error. Feedback in the form of rewards helps the system repeat desired behaviors and avoid inappropriate behaviors. This is a particularly powerful method for problems that require sequential decisions, such as robotics, logistics, and resource management.

However, a key problem with traditional RL approaches is that the learned strategy is often highly dependent on the training environment. Even small changes can cause an AI agent to no longer know how to behave properly. “Today’s RL agents perform admirably under trained conditions, but quickly reach their limits when conditions change,” Biedencap said. He will be working at the University of Freiburg until August 2026. From September 2026, he will lead the newly funded DFG Emmy Noether Group “From Mediocrity to Great Generalist: The Power of Context in RL” at KIT’s Institute of Anthropology. Robotics.

More context for a more robust learning process

Emmy Neether Group’s goal is to extend RL training methods so that AI becomes more robust and adaptive. To do so, Biedencup’s team will use additional information about the environment and world in which the agents operate. In this way, the AI ​​can learn which actions are most appropriate for which situations and later apply that knowledge to similar unknown situations.

In the long term, this approach could be an important step toward increasing the use of RL in real-world applications. Many RL-based AI systems to date have had to rely on highly accurate simulations of real-world environments, but the required simulators are complex, expensive, and difficult to implement in complex scenarios. “If RL-based systems can generalize better, it will no longer be so important to perfectly simulate every possible situation, which will greatly expand the range of possible applications of this technology,” Biedencup said.

About the Emmy Nether Program

The Emmy Noether Program gives highly talented scientists in the early stages of their careers the opportunity to lead their own research groups over six years and qualify for university professorship.

Detailed information