Towards high-level, verifiable autonomous behaviors with temporal specifications

We present a hybrid agent framework to produce high-level autonomous behavior for unmanned vehicles based on formal specification. High level autonomous behavior is attractive due to minimum level of central control and support for verifiable behavior results such as safety assurance. The proposed framework use linear temporal logic (LTL) to express high level agent behavior to control search, tracking, and survival activities, which are executed at a rule-based reasoning engine. The low level search and team tracking behaviors are implemented by a policy network trained with Reinforcement Learning (RL). The behavior controller is evaluated in a simulated envrionment with single-agent and multi-agent search and tracking scenarios.