Balancing Safety and Learning in the Internet of Things: A Risk-Aware Reinforcement Learning Approach

Reinforcement Learning (RL) enables Internet of Things (IoT) systems to adapt to dynamic environments, but ensuring safety during learning is challenging. Fallback-based safe reinforcement learning (SafeRL) methods guarantee safety by overriding unsafe actions, but hinder learning in high-risk yet informative states and may disrupt previously learned safe behaviors. We propose RA-SafeRL, a Risk-Aware SafeRL framework that improves learning without compromising safety. RA-SafeRL uses a digital twin environment to safely explore high-risk actions and learn from them, applies adaptive fallback penalties reflecting action severity to distinguish risk levels, and reuses fallback trajectories via off-policy learning, converting previously discarded data into useful signals. Experiments show RA-SafeRL outperforms baseline SafeRL by improving task performance by over 30%, reducing fallback frequency by more than 10%, and shortening recovery time by 38%, enhancing safety, efficiency, and robustness in safety-critical IoT applications.