Model-Free Motion Planning of Complex Tasks Subject to Ethical Constraints

Artificial Intelligence (AI) ethics establishes a moral framework to guide responsible AI technology use. This paper introduces a model-free Reinforcement Learning (RL) approach to address ethical constraints in motion planning problems, particularly in complex tasks within partially observable environments. Leveraging the Partially Observable Markov Decision Process (POMDP) for motion planning in environments with incomplete knowledge and Linear Temporal Logic (LTL) for task formulation, ethical norms are categorized as ‘hard’ and ‘soft’ constraints. Our approach involves generating a product of POMDP and LTL-induced automaton. An optimal policy is then learned, ensuring task completion while adhering to ethical constraints through model checking. To handle the situations where the agent lacks task awareness, we propose a novel modification to deep Q-learning. This model-free deep RL method employs a neural network architecture with environmental observations and recognized labels as inputs. An illustrative example showcases the applicability of our approach to motion planning problems. The flexibility and generality of this method make it suitable for addressing various ethical decision-making problems.