Adaro-RL

Robustness is a key requirement for building trustworthy RL systems. These systems are AI-based agents that interact with an environment, choosing actions over a sequence of states to achieve long-term goals. Robustness is crucial to ensure that an RL agent's behavior remains stable under diverse conditions and various types of perturbations. Adversarial attacks are one such perturbation: they introduce subtle changes to the agent's observations or to the dynamics of its environment.

Adaro-RL is a library of adversarial attacks and adversarial training methods dedicated to RL systems. It enables users to evaluate and train RL agents against diverse adversarial attacks, improving their robustness to observation perturbations and dynamics uncertainties.

It shall be used by ML scientist and engineer for RL agents development during the activity "Develop and acquire ML models" .