Reinforcement learning algorithms solve sequential decision-making problems in probabilistic environments by optimizing for long-term reward. The desire to use reinforcement learning in safety-critical settings inspires a recent line of work on formally constrained reinforcement learning; however, these methods place the implementation of the learning algorithm in their Trusted Computing Base. The crucial correctness property of these implementations is a guarantee that the learning algorithm converges to an optimal policy.

Cyber-physical systems, such as self-driving cars or autonomous aircraft, must defend against attacks that target sensor hardware. Analyzing system design can help engineers understand how a compromised sensor could impact the system's behavior; however, designing security analyses for cyber-physical systems is difficult due to their combination of discrete dynamics, continuous dynamics, and nondeterminism. This paper contributes a framework for modeling and analyzing sensor attacks on cyber-physical systems, using the formalism of hybrid programs. We formalize and analyze two relational properties of a system's robustness. These relational properties respectively express (1) whether a system's safety property can be influenced by sensor attacks, and (2) whether a system's high-integrity state can be affected by sensor attacks. We characterize these relational properties by defining an equivalence relation between a system under attack and the original unattacked system. That is, the system satisfies the robustness properties if executions of the attacked system are appropriately related to executions of the unattacked system. We present two techniques for reasoning about the equivalence relation and thus proving the relational properties for a system. One proof technique decomposes large proof obligations to smaller proof obligations. The other proof technique adapts the self-composition technique from the literature on secure information-flow, allowing us to reduce reasoning about the equivalence of two systems to reasoning about properties of a single system. This technique allows us to reuse existing tools for reasoning about properties of hybrid programs, but is challenging due to the combination of discrete dynamics, continuous dynamics, and nondeterminism. To validate the usefulness of our relational properties and proof techniques, we present three case studies motivated by real design flaws in existing cyber-physical systems.

Deploying deep reinforcement learning in safety-critical settings requires developing algorithms that obey hard constraints during exploration. This paper contributes a first approach toward enforcing formal safety constraints on end-to-end policies with visual inputs. Our approach draws on recent advances in object detection and automated reasoning for hybrid dynamical systems. The approach is evaluated on a novel benchmark that emphasizes the challenge of safely exploring in the presence of hard constraints. Our benchmark draws from several proposed problem sets for safe learning and includes problems that emphasize challenges such as reward signals that are not aligned with safety constraints. On each of these benchmark problems, our algorithm completely avoids unsafe behavior while remaining competitive at optimizing for as much reward as is safe. We characterize safety constraints in terms of a refinement relation on Markov decision processes - rather than directly constraining the reinforcement learning algorithm so that it only takes safe actions, we instead refine the environment so that only safe actions are defined in the environment's transition structure. This has pragmatic system design benefits and, more importantly, provides a clean conceptual setting in which we are able to prove important safety and efficiency properties. These allow us to transform the constrained optimization problem of acting safely in the original environment into an unconstrained optimization in a refined environment.

The desire to use reinforcement learning in safety-critical settings has inspired a recent interest in formal methods for learning algorithms. Existing formal methods for learning and optimization primarily consider the problem of constrained learning or constrained optimization. Given a single correct model and associated safety constraint, these approaches guarantee efficient learning while provably avoiding behaviors outside the safety constraint. Acting well given an accurate environmental model is an important pre-requisite for safe learning, but is ultimately insufficient for systems that operate in complex heterogeneous environments. This paper introduces verification-preserving model updates, the first approach toward obtaining formal safety guarantees for reinforcement learning in settings where multiple possible environmental models must be taken into account. Through a combination of inductive data and deductive proving with design-time model updates and runtime model falsification, we provide a first approach toward obtaining formal safety proofs for autonomous systems acting in heterogeneous environments.

Formal verification provides a high degree of confidence in safe system operation, but only if reality matches the verified model. Although a good model will be accurate most of the time, even the best models are incomplete. This is especially true in Cyber-Physical Systems because high-fidelity physical models of systems are expensive to develop and often intractable to verify. Conversely, reinforcement learning-based controllers are lauded for their flexibility in unmodeled environments, but do not provide guarantees of safe operation. This paper presents an approach for provably safe learning that provides the best of both worlds: the exploration and optimization capabilities of learning along with the safety guarantees of formal verification. Our main insight is that formal verification combined with verified runtime monitoring can ensure the safety of a learning agent. Verification results are preserved whenever learning agents limit exploration within the confounds of verified control choices as long as observed reality comports with the model used for off-line verification. When a model violation is detected, the agent abandons efficiency and instead attempts to learn a control strategy that guides the agent to a modeled portion of the state space. We prove that our approach toward incorporating knowledge about safe control into learning systems preserves safety guarantees, and demonstrate that we retain the empirical performance benefits provided by reinforcement learning. We also explore various points in the design space for these justified speculative controllers in a simple model of adaptive cruise control model for autonomous cars.

Autonomous cyber-physical systems-such as self-driving cars and autonomous drones-often leverage artificial intelligence and machine learning algorithms to act well in open environments. Although testing plays an important role in ensuring safety and robustness, modern autonomous systems have grown so complex that achieving safety via testing alone is intractable. Formal verification reduces this testing burden by ruling out large classes of errant behavior at design time. This paper reviews recent work toward developing formal methods for cyber-physical systems that use AI for planning and control by combining the rigor of formal proofs with the flexibility of reinforcement learning.

Differential dynamic logic is a logic for specifying and verifying safety, liveness, and other properties about models of cyber-physical systems. Theorem provers based on differential dynamic logic have been used to verify safety properties for models of self-driving cars and collision avoidance protocols for aircraft. Unfortunately, these theorem provers do not have explicit proof terms, which makes the implementation of a number of important features unnecessarily complicated without soundness-critical and extra-logical extensions to the theorem prover. Examples include: an unambiguous separation between proof checking and proof search, the ability to extract program traces corresponding to counter-examples, and synthesis of surely-live deterministic programs from liveness proofs for nondeterministic programs. This paper presents a differential dynamic logic with such an explicit representation of proofs. The resulting logic extends both the syntax and semantics of differential dynamic logic with proof terms -- syntactic representations of logical deductions. To support axiomatic theorem proving, the logic allows equivalence rewriting deep within formulas and supports both uniform renaming and uniform substitutions.