Towards provable probabilistic safety for scalable embodied AI systems