NeurIPS 2022

A Text-based Safety Benchmark for Reinforcement Learning Problems


In recent years, the issue of safety and robustness has become a critical focus for AI research. However, benchmarks for safe reinforcement learning tend to target a specific class of problems and do not offer a holistic set of challenges. In this paper, we propose a benchmark environment for safety-critical problems in deep reinforcement learning with text-based interaction. The contribution of this benchmark is in providing a general framework to incorporate safety constraints in agent interactions, as shown in our five problem gameset; moreover, the games can also be generated automatically to combine the multiple safety problems an agent might face. The source of safety constraints and goals are annotated from real-life examples of safety, and can be adapted to more open problems. Overall, our benchmark of Safety-critical Textworld is a flexible framework to provide a set of tasks to demonstrate a safety base challenges for reinforcement learning agents and aims to help the research community in exploring safety applications in a text-based domain.