Safe Planning with Artificial Intelligence 3.2. PROJECT SUMMARY (IN LAYMAN’S TERMS) (MAX. 300 WORDS) Artificial intelligence (AI) has entered our everyday life. With applications in areas as diverse as transportation, healthcare, and finance, the need for safety in AI has become of critical importance. We propose to develop automated methods for the safety assurance of AI systems. We focus on systems that operate under partial information and data uncertainty. Consider, for instance, a scenario where an autonomous agent needs to achieve some goal in an environment while avoiding collisions with other agents who have collaborative and conflicting goals. Such an agent may represent an autonomous car that has to account for other vehicles and pedestrians. Due to its restricted range of vision and sensors, the agent can only estimate the location and movements of the other agents - it has only partial information at its disposal. Moreover, the data concerning other environmental factors may be polluted or incomplete - resulting in uncertainty of the environment. The underlying problem is to compute a plan for the agent to reach its goal which accounts for such uncertainty and partial information - in real time. Such problems are very hard to solve and are often computationally intractable. We will investigate new approaches that exploit and augment methods from formal verification (FV), which subsumes a plethora of model-based techniques that provide guarantees on the correctness of systems. So far, the high complexity of AI systems has largely prevented an impact of FV on AI. We propose to bridge this gap by focussed research on the enabling factors uncertainty and partial information. We will (1) develop techniques that provide approximate yet sufficiently accurate solutions, (2) investigate natural restrictions of the problem such as limitations in memory or computation power of an agent.