Hsin-Jung Yang, Zhanhong Jiang, Prajwal Koirala, Qisai Liu, Cody Fleming, Soumik Sarkar
LexiSafe is a framework for offline safe reinforcement learning that prioritizes safety over rewards, reducing safety violations while improving task performance in cyber-physical systems.
LexiSafe is a new approach to offline reinforcement learning, which is particularly important for systems where safety is crucial, like self-driving cars or industrial robots. Traditional methods often struggle to maintain safety when trying to optimize for rewards, leading to potential safety violations. LexiSafe addresses this issue by prioritizing safety over rewards in its decision-making process, ensuring that the system behaves safely even when only pre-collected data is available for training. This approach not only reduces safety breaches but also enhances overall task performance, making it a promising solution for critical systems where safety cannot be compromised.
