Notions of Safety with Uncertainty in Specifications


Imagine the following tricky driving situations you face in everyday life: encountering a stopped vehicle or a cyclist on a two-lane road with opposing traffic flow. Although drivers are not allowed to cross the solid centerline to pass, it seems like a reasonable maneuver that will not compromise safety if performed correctly. In this case, the driver deviated from a rule while still maintaining safety. We posit that this is an instance of a problem where the specifications of the task are described as a distribution over a set of logical formulas. Satisfying all the formulas is sufficient but not all of them are necessary for safety. Thus the agent acting in the domain must try to best adhere to this distribution over formulas. We propose four evaluation criteria to capture the semantics of adhering to a distribution over specifications: most likely, maximum coverage, minimum regret and chance constrained. Next, using a simple synthetic task, we will demonstrate how the selection of the criterion results in qualitatively different executions depending on the nature of the belief distributions. We believe that this formulation will be particularly valuable in learning from demonstrations where there might be an inherent ambiguity in the task specifications. During the workshop we hope to spark discussion on methodology of selecting the right criterion, and utilization of these criteria in the design of exploration strategies that trade-off risk aversion and creativity.

Robotics: Science and Systems, Workshop on Robust Autonomy – safe robot learning and control in uncertain real-world environments (Poster)