Learning an interpretable logic monitor for risk-aware and socially-compliant trajectory planning

In order to plan robot behaviors effectively in the real world, it is often necessary to consider risks and learn from humans in dealing with them. We posit that humans manage risks by taking into consideration the nuances of the task that are specific to the current location and social context. We leverage past time signal temporal logic (ptSTL) formulas for forming compact, human-interpretable notions of risk. We introduce LogicRiskNet , a logic monitor constructed from ptSTL formulas that provide parameterized risk metrics and allow risk parameters to be learned from demonstration data capturing human behavior in risky situations. LogicRiskNet can be used to reason about environment agent behaviors and be incorporated into the controlled agent’s planner. To achieve human-like risk awareness, we explore an online adaptation mechanism that allows LogicRiskNet to update its parameters online in order to adhere to the aggregate behavior of its surrounding agents, while also benefitting from prior experience learned offline from large-scale datasets. We integrate LogicRiskNet in an inverse optimal control (IOC) framework and evaluate it on a real-world driving dataset. We show that our approach learns to generate trajectory plans that mimic the expert’s risk handling behaviors from offline demonstrations and adapts online to surrounding traffic at deployment time .