Provably Safe Reinforcement Learning via Action Projection Using Reachability Analysis and Polynomial Zonotopes

Abstract

While reinforcement learning produces very promising results for many applications, its main disadvantage is the lack of safety guarantees, which prevents its use in safety-critical systems. In this work, we address this issue by a safety shield for nonlinear continuous systems that solve reach-avoid tasks. Our safety shield prevents applying potentially unsafe actions from a reinforcement learning agent by projecting the proposed action to the closest safe action. This approach is called action projection and is implemented via mixed-integer optimization. The safety constraints for action projection are obtained by applying parameterized reachability analysis using polynomial zonotopes, which enables to accurately capture the nonlinear effects of the actions on the system. In contrast to other state-of-the-art approaches for action projection, our safety shield can efficiently handle input constraints and dynamic obstacles, eases incorporation of the spatial robot dimensions into the safety constraints, guarantees robust safety despite process noise and measurement errors, and is well suited for high-dimensional systems, as we demonstrate on several challenging benchmark systems.

Safe Reinforcement Learning for Urban Driving using Invariably Safe Braking Sets

“…This significantly increases the challenge of verifying safe actions because there are infinite individual continuous actions in an continuous action space. One approach could be obtaining rule-compliant state sets as proposed in [58] and correcting actions proposed by the agent to safe actions, e.g., with action projection as in [59].…”

Safe Reinforcement Learning for Automated Vehicles via Online Reachability Analysis

“…On the contrary, set-based predictions using reachability analysis [54] provide an over-approximation of all feasible future movements of other traffic participants that adhere to traffic rules; thus, safety can be guaranteed. Among all previous works, the closest to ours are [11], [42]. Kochdumper et al [42] propose a reachability-based safety shield for RL controllers for general nonlinear systems.…”

NNV 2.0: The Neural Network Verification Tool

“…For example, α,β-Crown was the top performer on last year's NN verification competition [55], able to verify FFNN, CNN and SSNNs, but it lacks support for neural ODEs or NNCS. There exist other tools that focus more on the verification of NNCS such as Verisig [34,35], Juliareach [63], ReachNN [17,33], Sherlock [16], RINO [26], VenMas [1], POLAR [32], and CORA [3,42]. However, their support is limited to NNCS with a linear, nonlinear ODE or hybrid automata as the plant model, and a FFNN as the controller.…”