Causal Liability in Autonomous Systems

With the widespread use of autonomous systems, liability of-<br/>ten shifts towards the manufacturers and part suppliers, especially when<br/>various system components, sourced from different suppliers, contribute<br/>to a failure. This necessitates a framework for automatic liability apportionment which can be embedded in manufacturer-supplier contracts and<br/>minimise legal disputes. To this end, we propose a formal framework based<br/>on the notion of actual causality in structural causal models and the robustness semantics of logical specifications. We prove several desirable<br/>properties of this framework. Moreover, we formalise the notion of causal<br/>non-interaction, give sufficient conditions for it to hold, and demonstrate<br/>its utility in deriving upper bounds for liability analysis. Furthermore, we<br/>relate our definition to the existing notion of harm, empirically evaluate<br/>our framework to demonstrate its efficacy, and release a software package<br/>implementing our approach. We extend our framework to handle interval<br/>specifications.