SafeEx: Hata Önleme Becerileri ile Güvenli Robot Yürütme Mimarisi

Ensuring safe robot execution in unstructured environments remains a significant challenge, particularly for learning-based robotic systems. In this work, we propose SafeEx, a safe robot execution framework that maintains task-oriented skills and safety-oriented failure prevention skills within a structured skill library. SafeEx represents failures using learned risk estimation models and enables efficient learning of failure prevention behaviors by generating risky states through aggressor skills for experimentation. During execution, a risk-based skill selection mechanism dynamically switches between base skills and prevention skills. SafeEx is evaluated in a robotic manipulation case study involving a stirring task with spill failures. Simulation results show that SafeEx reduces spill events by up to 87.5% compared to the baseline stirring policy. Overall, SafeEx significantly improves execution safety while maintaining task progression.