Training Verifiably Robust Agents Using Set-Based Reinforcement Learning
Manuel Wendl, Lukas Koller, Tobias Ladner, Matthias Althoff
TL;DR
This work integrates set-based reachability into reinforcement learning to train verifiably robust agents for continuous control. By propagating uncertainty sets through the actor and critic and optimizing a set-based loss, the method yields policies that minimize worst-case disturbances under an $\ell_\infty$ perturbation model and provides formal safety guarantees via reachability analysis. The authors derive explicit set-based losses and gradients, compare SA-SC and SA-PC against standard and adversarial baselines, and demonstrate robustness improvements across multiple benchmarks. The approach enables safer deployment of neural controllers in safety-critical settings and offers a pathway for rigorous verification in learning-based control systems.
Abstract
Reinforcement learning often uses neural networks to solve complex control tasks. However, neural networks are sensitive to input perturbations, which makes their deployment in safety-critical environments challenging. This work lifts recent results from formally verifying neural networks against such disturbances to reinforcement learning in continuous state and action spaces using reachability analysis. While previous work mainly focuses on adversarial attacks for robust reinforcement learning, we train neural networks utilizing entire sets of perturbed inputs and maximize the worst-case reward. The obtained agents are verifiably more robust than agents obtained by related work, making them more applicable in safety-critical environments. This is demonstrated with an extensive empirical evaluation of four different benchmarks.