Safety Considerations in Deep Control Policies with Safety Barrier Certificates Under Uncertainty

Recent advances in Deep Machine Learning have shown promise in solving\ncomplex perception and control loops via methods such as reinforcement and\nimitation learning. However, guaranteeing safety for such learned deep policies\nhas been a challenge due to issues such as partial observability and\ndifficulties in characterizing the behavior of the neural networks. While a lot\nof emphasis in safe learning has been placed during training, it is non-trivial\nto guarantee safety at deployment or test time. This paper extends how under\nmild assumptions, Safety Barrier Certificates can be used to guarantee safety\nwith deep control policies despite uncertainty arising due to perception and\nother latent variables. Specifically for scenarios where the dynamics are\nsmooth and uncertainty has a finite support, the proposed framework wraps\naround an existing deep control policy and generates safe actions by\ndynamically evaluating and modifying the policy from the embedded network. Our\nframework utilizes control barrier functions to create spaces of control\nactions that are safe under uncertainty, and when the original actions are\nfound to be in violation of the safety constraint, uses quadratic programming\nto minimally modify the original actions to ensure they lie in the safe set.\nRepresentations of the environment are built through Euclidean signed distance\nfields that are then used to infer the safety of actions and to guarantee\nforward invariance. We implement this method in simulation in a drone-racing\nenvironment and show that our method results in safer actions compared to a\nbaseline that only relies on imitation learning to generate control actions.\n