A survey on probabilistic planning and temporal scheduling with safety guarantees

Autonomous robotic systems are often safety-critical since unsafe behavior can have disastrous consequences. In this paper, we survey existing frameworks that can incorporate safety guarantees or constraints in the design of an autonomous system. Rather than verifying the guarantees during simulation or testing, such frameworks can enforce them upfront. Furthermore, in a physical setting, the effects of an agent's actions can not be considered deterministic. The frameworks have to take into account the uncertain or probabilistic effects of actions. Different frameworks provide different types of guarantees. Yet no comprehensive overview of such frameworks and the guarantees they offer exists today. This survey tries to answer the need for an overview. Probabilistic planning is often tackled using Markov Decision Processes (MDPs). Many extensions of MDPs exist, some of which can provide explicit safety guarantees. In the existing research, constraints on objective functions, reachable states, and execution paths of the system are obtained. For scheduling, Simple Temporal Networks (STNs) are addressed. STNs inherently incorporate temporal constraints, enforcing temporal relations. As an extension, probabilistic constraints on failure can be imposed as well.