Towards an Explainable Reinforcement Learning Agent for Resource-Constrained Robots

This paper considers the problem of how a robot during a planetary exploration mission can operate under strict resource constraints, requiring high transparency in decision-making: a context where failure is not an option. Standard reinforcement learning methods are often unsuitable for such environments due to their lack of transparency and high data requirements. In this work, we introduce SPARCbandit, a lightweight, explainable reinforcement learning agent that integrates linear Q-learning with tile coding, adaptive action cost tracking, and behavioural latching. We benchmark SPARCbandit in both standard control (LunarLander-v3) and simulated planetary locomotion (a quadruped lunar rover in PyBullet), demonstrating improved learning speed, stability, and transparency.